Robotics and Computer–Integrated Manufacturing 88 (2024) 102733

Contents lists available at ScienceDirect

Robotics and Computer-Integrated Manufacturing

journal homepage: www.elsevier.com/locate/rcim

Full length Article

Precision robotic deburring with Simultaneous Registration and Machining

for improved accuracy, quality, and efficiency
Steffan Lloyd ∗, Rishad A. Irani, Mojtaba Ahmadi
Department of Mechanical and Aerospace Engineering at Carleton University, Ottawa, Ontario, K1S 5B6, Canada

ARTICLE INFO ABSTRACT

Keywords: This paper extends and validates our recently developed Simultaneous Registration and Machining (SRAM)
Robotic deburring framework for robotic deburring applications. SRAM enhances contact stability and machining quality in planar
Admittance control robotic contouring tasks with uncertain registration. The framework estimates workpiece-tool contact points
Intelligent manufacturing systems
and utilizes them to improve registration accuracy and reduce path error in real-time, while modulating force
Adaptive control
controller tuning for optimal edge-finding ability or contact stability as required. We propose SRAM-D, an
enhanced deburring-specific SRAM controller that addresses practical shortcomings in the original framework
for improved workpiece quality and contact stability. Added virtual spring modulation ensures more effective
burr removal, while a novel feedrate modulation strategy improves contact stability and optimizes cycle times.
SRAM-D is implemented in a custom robotic deburring workcell and validated in extensive experimental testing
against nominal SRAM and nominal admittance (ADMIT) controllers with varying registration accuracy. Tests
demonstrate that SRAM-D outperforms both nominal controllers in all tested metrics. Path error is minimized
and contact stability is improved, enhancing workpiece quality. SRAM-D improves large burr removal by 34%
over the nominal SRAM and 45% over ADMIT. Furthermore, SRAM-D reduces tool wear by 29% compared
to ADMIT, significantly decreasing ongoing tool costs. SRAM-D improves the feasibility of precision robotic
deburring in industry by enhancing workpiece quality while enabling the use of lower accuracy yet faster and
less costly registration methods—ultimately reducing workcell costs and boosting process efficiency.

1. Introduction

Deburring is a secondary operation required after many metal-

forming processes. Punching and shearing produce tear burrs, shown
in Fig. 1a, while nibbling, which involve overlapping die punches,
creates sharp scallops that require removal, shown in Fig. 1b. Nibbling
also leaves index tabs—an intentional scallop to hold the workpiece
during operations, as in Fig. 1c. These imperfections can cut a worker’s
hand, damage electrical cables, prevent parts from mating properly,
cause early corrosion, or allow for crack initiation and lead to early
failure [1]. In the aerospace industry, strict tolerances, low produc- Fig. 1. (a) Examples of tear burrs left after punching, (b) examples of nibbling scallops,
(c) example of an index tab.
tion volume, and varied workpiece sizes render current automation
systems unsuitable, and manual hand deburring is widely employed
instead [2,3]. Hand deburring is also prevalent in the automotive sector
because of high part complexity [4]. Unfortunately, manual deburring higher percentages, from 10% to 35% [3,8]. Thus, it is highly desirable
is costly and error-prone [1]. It is associated with high worker turnover to automate this process and allow for more flexible and scalable
and respiratory and vibration-induced health risks [5]. Consequently, deburring in manufacturing.
manual deburring is challenging to scale and is often identified as a Industrial robots present an attractive deburring automation so-
bottleneck in industry [6]. One study reported that deburring com- lution as they are cost-effective, scalable, reconfigurable, and have
prised 9% of their total manufacturing costs, [7], while others report

∗ Corresponding author.
E-mail addresses: [email protected] (S. Lloyd), [email protected] (R.A. Irani), [email protected] (M. Ahmadi).

https://doi.org/10.1016/j.rcim.2024.102733
Received 6 July 2023; Received in revised form 22 December 2023; Accepted 26 January 2024
Available online 1 February 2024
0736-5845/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

the tool and workpiece. In some cases, contact stability has been
achieved using passive-compliance deburring end-effectors [36]; how-
ever, these devices have limited accuracy and are unsuitable when tol-
erances are tight [37]. For higher-precision applications, force-control
methods are employed instead. Several different force control strategies
have been investigated for robotic deburring. Pure force control has
stability issues when implemented on hardware, so force control is
typically implemented using interaction controllers. These controllers
define a target dynamic relationship between force and position—
usually in the form of a mass–damper–spring system that forms a
‘‘virtual suspension’’ between the tool and the part.
Impedance control [38] has been investigated with simplified one-
or two-degree-of-freedom (DOF) manipulators [39–41], but is challeng-
ing to implement on industrial robots, as accurate control requires
full modeling and compensation of the complex manipulator dynamics
and nonlinear friction forces [42]. Admittance control [43] is more
applicable to articulated robotic arms and has been used in literature
for deburring [25,28,37,44]. In [37], admittance-controlled deburring
was accomplished by performing two contouring passes of the part—
Fig. 2. Deburring contouring path (red) about a workpiece. the first to localize and the second to deburr. In [28], aluminum
castings were deburred using admittance control and touch-off local-
ization; however, accuracy and machining quality were not quantified.
In [44], the authors developed an admittance controller integrating
large, dexterous workspaces. However, robotic deburring has only been
sliding mode control for surface conditioning operations. Several re-
adopted in few, specific situations, and machining operations account
searchers have also investigated fuzzy control for deburring [45,46],
for only 0.6% of global robot usage [9]. Often cited factors to this
which implements controller behavior similar to admittance control
limited adoption include insufficient accuracy, low structural rigidity
using fuzzy logic rules, which can be easier to interpret and tune in
leading to vibrations or chatter, and challenges registering the work-
some situations. Unfortunately, the proposed strategies have overall not
piece in the robot’s workspace [10–13]. Thus, successful robotization of
advanced sufficiently for practical industrial implementation of robotic
the deburring process requires innovative approaches to improving the
deburring. Accuracy and machining quality are often poor or not re-
accuracy of the manipulator and the workpiece registration, while min-
ported. Contact stability is rarely quantified, and workpiece registration
imizing vibrations and simplifying operations. In this paper, we adapt,
improve, and validate our recently introduced control framework, Si- is typically only addressed through prohibitively slow methods. Tool
multaneous Registration and Machining (SRAM) [14,15] to address these wear concerns have not been properly investigated nor quantified in
robotic deburring challenges. The SRAM framework refines the pla- the deburring context.
nar workpiece registration in real-time by leveraging measured force Moreover, the admittance control framework often utilized in liter-
and position data. Simultaneously, a force control algorithm corrects ature has fundamental limitations when applied to robotic deburring.
remaining path error, with adaptive tuning that minimizes vibrations An admittance controller is tuned with virtual mass, spring, and damp-
and optimizes surface finish. The result is improved accuracy and ing coefficients, where the damping coefficient in particular changes
contact stability throughout the process. controller behavior. When tuned with low damping, an admittance con-
Fig. 2 illustrates a typical robotic deburring operation, with a planar troller can move quickly, behaving similarly to a pure force controller
workpiece mounted to a vacuum clamp and a contouring path for the with ideal edge-tracking. However, limited robot bandwidth results in
manipulator to follow. Process accuracy depends on two main factors: chatter when damping is low, giving poor surface finish, uneven de-
the manipulator accuracy and the workpiece registration quality. While burring, and premature tool wear [15,47]. Alternatively, the controller
the former can be improved with established, albeit complex, kinematic can be tuned with high damping to behave similarly to a pure position
calibration procedures [16–19], workpiece registration is challenging controller, which minimizes chatter. Unfortunately, this high-damping
since it must be completed on-the-fly for each new workpiece. Various configuration cannot quickly compensate for path errors, causing the
methods have been explored for registration, each with drawbacks. tool to damage the part or lose contact with the workpiece, leav-
Precision jigging is possible but is costly and unsuitable when work- ing unprocessed edges. Other aspects of deburring further compound
piece variability is high [20]. Laser scanners can be used for accurate these challenges. The workpiece-tool contact is typically very rigid,
localization [16,19,21–24] but are expensive, and typically only pro- aggravating chatter issues. In addition, the high-speed rotation of the
vide 2D information. Thus, the scanner must be physically moved deburring spindle can impart large forces and destabilize workpiece-
over the part to generate a point cloud for registration, significantly tool interactions. Lastly, as deburring tools wear in the contouring path,
slowing cycle times. 3D laser scanners exist but have limited resolu- their radius can change and introduce additional path errors that must
tion and very high cost [22,23]. Laser systems can also be affected be compensated for.
by workpiece reflectivity or skewed by the presence of large burrs Nominally, admittance controllers are also reactive to part geometry
or flashing. Touch-off testing can be used to localize parts [20,25– and do not gain any fundamental intuition about the location of the
28] but is also prohibitively slow since robot speed is limited near part. As the tool contours the part, every workpiece corner introduces
unknown geometries. Moreover, the part can only be sampled sparsely, a different amount of path error for the force controller to compensate
making touch-off testing vulnerable to outliers from burrs. Lastly, vision for, since error along one edge is uncorrelated to error along a perpen-
systems can be used [29–33], and are cost-effective and report results dicular edge. Thus, every corner around the entire workpiece periphery
quickly [10]. However, vision systems have limited accuracy [34,35] requires new adaptation, even after the system has ‘‘observed’’ the
and are also sensitive to visual outliers from burrs. Issues can also workpiece sufficiently to learn its true location, giving sub-optimal
arise during the object detection and image processing steps, reducing results.
registration reliability and accuracy. Recently, we introduced a novel framework called Simultaneous
When workpiece registration is imperfect, additional steps are nec- Registration and Machining, or SRAM (see Lloyd et al. [14,15]). The
essary to ensure the contouring path maintains a stable contact between SRAM method aims to allow for less accurate workpiece registration

2
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

in planar robotic contouring applications while improving contact sta- attainable by leveraging the SRAM-D framework to compensate for the
bility, accuracy, and machining quality. SRAM uses acquired position increased error of faster, yet less accurate localization techniques.
and force data from the contouring of the workpiece to iteratively This work is organized as follows. Section 2 provides a review of the
improve registration and correct the machining path in real-time. Tool original SRAM framework, and introduces the proposed modifications
size and wear are modeled and compensated for online. Simultane- for improved burr removal and feedrate modulation. Section 3 intro-
ously, the admittance force controller is adaptively tuned based on live duces the experimental platform developed for testing, describes proce-
registration uncertainty—giving low damping and fast response when dures for calibration of the tool models used in the SRAM method, and
edge location is uncertain, but transitioning to a highly damped, slow provides detailed guidelines to tune and optimize the SRAM method.
response when possible to reduce chatter and optimize surface finish. Finally, Section 4 summarizes the experimental validation results in
The SRAM framework has been validated in preliminary testing and terms of path accuracy, contact stability, scallop removal, tool wear,
and cycle times.
shown to significantly reduce path errors in contouring applications,
give optimal contact stability, and lower tool wear compared to a
2. Simultaneous Registration and Machining for robotic deburring
nominal admittance controller.
SRAM appears to provide an ideal solution to many of the current This section provides a brief overview of the original SRAM method
challenges in precision robotic deburring. The contact stability tuning and introduces our proposed modifications and improvements for re-
trade-off is resolved optimally, reducing chatter and ensuring high liable precision robotic deburring. Section 2.1 introduces the prob-
machining quality. Accuracy and contact stability can be improved, lem statement, key concepts, and notation. Section 2.2 provides an
and by allowing for less accurate workpiece registration, faster and less overview of the original SRAM framework, as previously introduced
expensive localization methods such as vision systems can be employed, in [14,15]. Finally, Section 2.3 introduces improvements to adapt
improving cycle times and workcell efficiency. However, further work SRAM for precise robotic deburring.
is needed. In deburring, the workpiece edge must be smoothed, and
it is necessary to remove any burrs, scallops, or other workpiece 2.1. Problem overview & preliminaries
imperfections along the edge. In the original SRAM framework, a
constant workpiece-tool contact force was applied while contouring The SRAM framework addresses the problem of contouring a planar
the part, which may smooth the edge but not necessarily remove workpiece using force control, with possibly poor initial registration
larger edge imperfections. The original SRAM method also employed and an uncertain estimate of the tool radius. SRAM improves con-
a constant-feedrate approach; however, modulating the feedrate in touring performance in these scenarios by correcting for path error
real-time could remove burrs more effectively while also minimizing using real-time acquired position and force data, and adaptively tuning
the admittance force controller for optimal performance based on
cycle times. Lastly, the SRAM method has only undergone limited
workpiece placement covariance. The described setup was visualized in
preliminary testing in contouring applications. More comprehensive,
Fig. 2. We define 𝐓𝑝 as an initial, potentially inaccurate estimate of the
statistically significant results in deburring are needed to fully char-
homogeneous transform representing the workpiece pose, where the 𝑥-
acterize the advantages of the method, and validate it for industrial
𝑦 plane coincides with the plane of the part. The symbol 𝑟𝑡 similarly
adoption.
denotes an initial, potentially poor estimate of the tool radius.
This paper aims to adapt, improve, and validate the recently intro- The part’s geometry is assumed to be available as 2D primitives such
duced SRAM method for precision robotic deburring applications. We as lines and arcs, and it is assumed that a path can be planned around
introduce an enhanced SRAM method for robotic deburring (SRAM- the part based on the initial pose and tool radius, as shown in Fig. 2.
D), which builds on the original SRAM framework by introducing two This pre-planned path contains three parameters: 𝐩𝑝 , the Cartesian
major improvements: coordinates of the robot tool, 𝐮𝑛 , a unit vector pointing along the
direction normal to the workpiece edge during contouring, and 𝑓𝑑 , the
1. First, as SRAM enables real-time correction of the contouring
desired normal workpiece-tool contact force. Planning such a path can
path, path error is quickly reduced to negligible levels during
be achieved using methods described in various published works [26,
the run. This improved accuracy can be leveraged to identify
27,29,48,49], or using off-the-shelf CAD/CAM software packages.
and fully remove large burrs, scallops, or index tabs on the part, For the following analysis, it is helpful to use planar coordinates
rather than merely passing over them with constant force. We within the 𝑥–𝑦 plane of 𝐓𝑝 . To differentiate the frames in this paper,
propose a modification to the SRAM admittance adaptation laws we use the underbar □ to distinguish 2D planar variables from their 3D
to penalize positional deflections when part registration is good, ̄
counterparts. Converting a 3D point 𝐩 to its planar counterpart 𝐩 (and
allowing for more complete removal of burrs. vice-versa) is accomplished with the unit vectors 𝐮𝑥𝑝 and 𝐮𝑦𝑝 alonḡ the
2. Second, we introduce a novel feedrate modulation technique to 𝑥 and 𝑦 axes of the part frame 𝐓𝑝 as
more effectively remove large workpiece imperfections, improve [ ]𝖳 ( ) [ ]
controller contact stability, and optimize cycle times. 𝐩 = 𝐮𝑥𝑝 𝐮𝑦𝑝 𝐩 − 𝐝𝑝 , 𝐩 = 𝐮𝑥𝑝 𝐮𝑦𝑝 𝐩 + 𝐝𝑝 . (1)
̄ ̄
Additionally, this paper offers several other contributions. We out-
2.2. Simultaneous Registration and Machining
line a comprehensive calibration procedure for the SRAM tool models,
and provide detailed guidelines to implement and tune the controller
This section provides an overview of our previously introduced
for optimal performance. The proposed SRAM-D controller is validated
SRAM framework, as in [15]. Fig. 3 presents the control architecture of
through comprehensive testing, with large sample size for robust and
the method. In SRAM, the robotic manipulator follows the pre-planned
statistically significant results, against the original SRAM method and path consisting of 𝐩𝑝 , 𝐮𝑛 , and 𝑓𝑑 . Simultaneously, an admittance con-
a nominal admittance controller (ADMIT) under various degrees of troller compensates for path errors by modulating a path correction 𝑎
registration uncertainty. The results show the superiority of the pro- along the normal axis 𝐮𝑛 . This corrected path is run through the inverse
posed controller in terms of increased path accuracy, enhanced contact kinematics algorithm to determine the correct joint angles 𝐪cmd , which
stability, and improved workpiece quality. SRAM-D is also shown to are then sent directly to the PID setpoints of the robot controller.
remove burrs and scallops more effectively than the nominal controllers While contouring the part, the SRAM Adaptation block monitors
and significantly reduce tool wear. Deburring quality of the tested the measured robot position 𝐪 and tool forces 𝑓 and attempts to
controllers is compared with manually deburred workpieces from a hu- improve the registration estimate of the part while adaptively tuning
man worker. Finally, we quantify cycle times for different registration the admittance controller for optimal performance. The system outputs
methods, revealing the potential improvements in workcell efficiency four key terms:

3
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Fig. 3. Control architecture of the SRAM framework.

2.2.1. Data collection

The SRAM system monitors the robot joint angles 𝐪 and measured
normal force, 𝑓 . The measured tool center 𝐩𝑡 is computed from the
measured joint angles with the forward kinematic ̄ model of the ma-
nipulator. At a given timestep, if the normal force exceeds a threshold
𝑓thresh , the algorithm classifies this point as ‘‘in contact’’ with the part.
The in-contact observations are added to a growing set of points 𝐩𝑡𝑖 and
𝑓𝑖 , which will be used to refine the workpiece location and tool̄ size.

2.2.2. Tool wear compensation

The second step uses a wear model to estimate changes in tool
radius based on the observed contact forces 𝑓𝑖 , as
Fig. 4. Nomenclature of the SRAM workpiece-tool interactions. 𝑡𝑖 ( ∞ )𝑗𝑡
𝛿𝑟𝑡,wear = 𝜓̃ 𝑡 𝛷𝑡𝑖 , 𝛷𝑡𝑖 ≜ 𝑓 (𝑡) 𝑑𝑡, (4)
∫0 0

where 𝜓̃ 𝑡 is the estimated tool wear rate and 𝛷𝑡𝑖 is the integral of
1. 𝐝cor , a path correction term based on an improved estimate of the exponentiated force to the 𝑖th time step. The shape parameter 𝑗𝑡
the workpiece location and tool size, captures the effect of higher forces giving exponentially higher wear,
2. 𝐑cor , a rotational correction term for the orientation of the nor- and the notation denotes a saturation between 0 and ∞, as negative
mal vector 𝐮𝑛 , based on an improved estimate of the workpiece forces do not result in negative tool wear. The force 𝑓 (𝑡) must also be
rotation, lowpass-filtered with a cutoff frequency 𝐹𝑡 to remove high-frequency
3. 𝛾𝑎 , an adaptive tuning factor in the range 0–1, which represents
noise. The tool model calibration and selection of 𝐹𝑡 is discussed further
the quality of the algorithm’s estimate of the workpiece location,
in Section 3.2.
projected onto the current normal axis 𝐮𝑛 . When 𝛾 → 1, the
uncertainty of the path correction 𝑎 is high, and when 𝛾 → 0,
the uncertainty is low. 2.2.3. Estimation of workpiece-tool contact points
4. 𝑎̇ cor , a velocity correction term added in the admittance con- The third step uses the measured tool centers 𝐩𝑡𝑖 and forces 𝑓𝑖 to
troller to compensate for changes in the path correction terms ̄
estimate the workpiece-tool contact points 𝐩𝑐𝑖 , by translating from 𝐩𝑡𝑖
𝐝cor and 𝐑cor . along the normal axis as ̄ ̄

The admittance controller defines the dynamic relationship between 𝐩̃ 𝑐𝑖 = 𝐩̃ 𝑡𝑖 + 𝑟𝑡𝑖,cor 𝐮̃ 𝑛𝑖 . (5)
positional and force error to maintain the desired contact force 𝑓𝑑 by ̄ ̄ ̄
modifying the commanded tool center a distance 𝑎 along the corrected Here, 𝑟𝑡𝑖,cor is the corrected tool radius for the 𝑖th sample,
normal axis 𝐮̃ 𝑛,cor , as depicted in Fig. 4, as
𝑓𝑖
𝑟𝑡𝑖,cor = 𝑟𝑡 + 𝛿𝑟𝑡 − 𝜓̃ 𝑡 𝛷𝑡𝑖 − , (6)
𝐩cmd = 𝐩𝑝 + 𝑎 𝐮̃ 𝑛,cor + 𝐝cor , 𝐮̃ 𝑛,cor = 𝐑cor 𝐮𝑛 , (2) 𝑘𝑡

where 𝑎 is computed with mass 𝑚𝑎 , spring constant 𝑘𝑎 , and damping which incorporates the initial radius correction term 𝛿𝑟𝑡 , the wear cor-
dynamically modulated as 𝑏𝑎 ∕𝛾𝑎 , such that rection term 𝜓̃ 𝑡 𝛷𝑡𝑖 , and a spring term accounting for the tool deflection
[ ] ( ) due to the applied force 𝑓𝑖 , with an equivalent workpiece-tool contact
1 𝑏
𝑎̈ = 𝑓𝑑 − 𝑓 − 𝑎 𝑎̇ − 𝑘𝑎 𝑎 , 𝑎= 𝑎̇ cor + 𝑎̈ 𝑑𝑡 𝑑𝑡. (3) stiffness, 𝑘𝑡 . A procedure to calibrate 𝑘𝑡 is presented in Section 3.2.
𝑚𝑎 𝛾𝑎 ∫ ∫
When the part registration is poor, 𝛾𝑎 ≈ 1, the effective damping is
low, and the controller can move quickly to locate workpiece edges. 2.2.4. Matching with CAD geometry
However, as the localization improves, 𝛾𝑎 → 0, the damping increases, The next step in the SRAM algorithm involves matching the esti-
and the controller is slowed for better contact stability. mated contact points with the CAD geometry of the estimated work-
The SRAM Adaptation system internally performs several steps to piece by computing the point 𝐩𝑚𝑖 , on the CAD geometry, that is geomet-
generate its outputs. A detailed understanding of these steps is not ̄
rically closest to each of the contact points, 𝐩̃ 𝑐𝑖 . The equations to locate
crucial for this work, but can aid in comprehending the method’s ̄
the matched points 𝐩𝑚𝑖 differ by primitive (line, arc, spline, point, etc.)
tuning parameters. We provide a high-level overview of the algorithm ̄
and parameterizations, and is described fully in [15]. After identifying
steps, focusing on key concepts and tuning parameters. Other details 𝐩𝑚𝑖 , the residuals are defined as
are omitted for brevity and can be found in [15]. The computations ̄
comprise nine steps, as follows. 𝐞𝑖 = 𝐩𝑐𝑖 − 𝐩𝑚𝑖 . (7)
̄ ̄ ̄
4
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

2.2.5. Detection of outliers and residual weighting 2.2.7. Correction of the contouring path
The contact model in Section 2.2.3 will not be perfect. Workpiece As the state 𝛒 is updated in real-time, improved information about
imperfections and complex workpiece geometry can cause the tool the correct contouring path becomes available. This information is
to contact the workpiece unexpectedly, resulting in anomalous data. reintroduced into the algorithm via the correction variables 𝐝cor and
A two-stage outlier detection method is used to prevent erroneous 𝐑cor as in Fig. 3 as
estimated contact points from skewing the algorithm. First, a distance-
𝐩𝑝,cor = 𝐩𝑝 + 𝐝cor , 𝐮𝑛,cor = 𝐑cor 𝐮𝑛 . (14)
based weighting scheme fits a multivariate Gaussian distribution to
the residuals 𝐞𝑖 , assigning a weight 𝑤𝑑𝑖 to each residual based on the The 2D vector 𝐝cor comprises the translational correction 𝛿 𝐝 =
̄ [ ] ̄ ̄
Mahalanobis distance 𝑑𝑚𝑖 (i.e., statistical distance from the mean) as 𝛿𝑥 𝛿𝑦 𝖳 , a rotation 𝜃, and a radius-correcting translation along 𝐮̃ 𝑛 ,
( )−1 ̄
( ) ( )
𝑤𝑑𝑖 = 1 + 𝑒4(𝑑𝑚𝑖 −𝑏𝑑 )∕𝑎𝑑 . (8) 𝐝cor = 𝛿 𝐝 + 𝐑𝜃 𝐩𝑝 − 𝐩𝑝 + 𝑟𝑡 − 𝑟𝑡,cor 𝐑𝜃 𝐮𝑛 , (15)
̄ ̄ ̄ ̄ ̄ ̄ ̄
Eq. (8) defines a sigmoidal curve that returns a value near one when 𝑑𝑚𝑖 where 𝐑𝜃 is the 2 × 2 rotation matrix for an angle 𝜃. The 3D vector 𝐝cor
̄
is small, and transitions to zero as 𝑑𝑚𝑖 increases. The tuning parameters is obtained by projecting the planar 𝐝cor into 3D space as in Eq. (1).
̄
𝑎𝑑 and 𝑏𝑑 control the transition shape—𝑎𝑑 sets the transition steepness, Similarly, 𝐑cor is defined as the 3D projection of 𝐑𝜃 . As these path
such that 96% of the threshold transition occurs within 𝑎𝑑 standard adjustments change the admittance correction distance 𝑎 required by
deviations, while 𝑏𝑑 sets the transition center. the force controller, a velocity correction term is 𝑎̇ cor is introduced
Second, a time-based weighting scheme protects newly acquired as in Eq. (3). This term compensates for effective changes in 𝑎 due to
points from the anomaly detection, as they can contain new informa- corrections in the path, and can be computed as
tion and be unweighted by Eq. (8). A second weight 𝑤𝑡𝑖 is defined based ( )𝖳 ( 𝜕𝛒 )
𝜕𝑎 𝜕𝑎
on sample age 𝛿𝑡𝑖 as 𝑎̇ cor = = = ∇𝑎𝖳 𝛒,
̇ (16)
𝜕𝑡 𝜕𝛒 𝜕𝑡
( )−1
where the gradient term ∇𝑎 is computed as
𝑤𝑡𝑖 = 1 + 𝑒4(𝛿𝑡𝑖 −𝑏𝑡 )∕𝑎𝑡 , (9)
[ ( ′ )𝖳 ( ) ]𝖳
∇𝑎 = 𝐮̃ 𝑛 𝖳 𝐑𝜃 𝐮𝑛 𝛿 𝐝 − 𝐩𝑝 −1 𝛷𝑡𝑘 . (17)
assigning a weight near one for new samples and transitioning to zero ̄ ̄ ̄ ̄ ̄
as sample age increases. As before, 𝑎𝑡 and 𝑏𝑡 control the steepness and
2.2.8. Estimation of the state covariance
center of the transition.
The state covariance 𝐏̃ provides a measure of the algorithm con-
The two weights, 𝑤𝑑𝑖 and 𝑤𝑡𝑖 , are then combined into a single
vergence and is used to adaptively tune the admittance controller.
weight 𝑤𝑖 using probabilistic ‘‘OR’’ logic, 𝑤𝑖 = 1 − (1 − 𝑤𝑑𝑖 ) (1 − 𝑤𝑡𝑖 ),
The state covariance is computed by linearly projecting the observa-
such that 𝑤𝑖 is large if either 𝑤𝑑𝑖 or 𝑤𝑡𝑖 are large, i.e., if the point is
tion covariances into state-space via the implicit function theorem [50],
not an outlier or if it is a newly acquired point.
resulting in the simplified closed-form equation [15]
[ ∑𝑛 ]
2.2.6. Iterative refinement of the workpiece registration 𝐏̃ = 𝐇−1 𝐏−1
0
+ 𝑤𝑖 2 ∇𝐞𝑖𝖳 𝐒−1 −1
𝑖 ∇𝐞𝑖 𝐇 . (18)
𝑖=1
̄ ̄
The sixth step uses the residuals 𝐞𝑖 and weights 𝑤𝑖 to refine the
̄ 2.2.9. Adaptive tuning of the admittance force controller
workpiece registration. The system uncertainty is packaged into a
five-element state vector 𝛒 as The final step in the SRAM adaptation computes the adaptive tuning
[ ]𝖳 term 𝛾𝑎 that modulates the admittance virtual damping as in Eq. (3).
𝛒 = 𝛿𝑥 𝛿𝑦 𝜃 𝛿𝑟𝑡 𝛿𝜓𝑡 , (10) First, the state covariance 𝐏̃ is projected onto the admittance axis to
where 𝛿𝑥, 𝛿𝑦, and 𝜃 are translational and rotational corrections in the obtain the admittance variance as
workpiece plane, 𝛿𝑟𝑡 is a correction to the tool radius, and 𝛿𝜓𝑡 is a 𝜎𝑎2 = var(𝑎) = ∇𝑎𝖳 𝐏̃ ∇𝑎. (19)
correction to the initial wear rate estimate such that 𝜓̃ 𝑡 = 𝜓𝑡 + 𝛿𝜓𝑡 .
Refinement of 𝛒 is accomplished by minimizing the quadratic cost The admittance variance 𝜎𝑎2
quantifies the local variance in the admit-
function tance correction, and provides an accurate metric to determine when
the controller requires low damping for better edge-finding ability, or
∑
𝑛
𝐿(𝛒) = 12 𝛒𝖳 𝐏−1
0
𝛒+ 1
2
𝑤𝑖 𝐞𝑖𝖳 𝐒−1
𝑖 𝐞𝑖 , (11) high damping for better contact stability. The admittance adaptation
𝑖=1
̄ ̄ parameter 𝛾𝑎 is computed by transforming 𝜎𝑎 using the heuristic
( )1
where 𝐏0 is a 5 × 5 positive definite initial state covariance matrix, 1 − (𝜎𝑎0 ∕𝜎𝑎 )𝑗𝑎
describing the accuracy of the poor initial registration estimate, and 𝛾𝑎 = 1 − 𝑗
(20)
1 − (𝜎𝑎0 ∕𝜎𝑎,min ) 𝑎
𝐒𝑖 is a 2 × 2 positive definite matrix describing the covariance of the 𝛾𝑎,min

𝑖th measurement. 𝐿(𝛒) is iteratively refined using a damped Newton where 𝜎𝑎0 is an estimate of the initial admittance standard error, 𝜎𝑎,min
update, as is an estimate of the minimum admittance standard error associated
[ ]−1 with perfect workpiece registration, 𝑗𝑎 is a tunable shape parameter,
𝛒𝑘+1 = 𝛒𝑘 − 𝐇(𝛒𝑘 ) + diag 𝛌 𝐠(𝛒𝑘 ), (12) and 𝛾𝑎,min is used to saturate 𝛾𝑎 within an acceptable and stable range,
where 𝐠(𝛒𝑘 ) ∈ R5 and 𝐇(𝛒𝑘 ) ∈ R5×5 are the optimization gradient and ensuring controller stability. Tuning of these parameters is discussed
Hessian, respectively, computed as further in Section 3.3.
∑
𝐠(𝛒) = 𝐏−1
0
𝛒+ 𝑤𝑖 ∇𝐞𝑖𝖳 𝐒−1
𝑖 𝐞𝑖 , 2.3. SRAM-D: Simultaneous Registration and Machining for deburring ap-
∑𝑖 [ ̄ ̄ ]
𝐇(𝛒) = 𝐏−1 + ∇ 𝖳 −1 2 −1
(13) plications
0
𝑤 𝑖 𝐞 𝑖 𝐒𝑖 ∇𝐞𝑖 + ∇ 𝐞𝑖 𝐒𝑖 𝐞𝑖 ,
𝑖 ̄ ̄ ̄ ̄
and 𝛌 ∈ R5 is a vector of damping terms for each state variable. The The preceding section provided an overview of the previously-
gradient and Hessian terms ∇𝐞𝑖 and ∇2 𝐞𝑖 can be analytically derived developed SRAM method for general contouring applications. This
̄ ̄
and computed [15]. The tunable damping vector 𝛌 moderates the section introduces the enhanced deburring-specific SRAM-D controller,
convergence of the algorithm to prevent updates from being applied which addresses practical shortcomings in the original framework for
faster than the robot can respond. It can also be employed to prioritize improved workpiece quality and contact stability. Two novel improve-
convergence of certain state variables over others, when states are not ments are proposed in the following subsections. The first increases
individually observable. deburring accuracy by more effectively leveraging the improved path

5
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

estimate of the SRAM method to remove large burrs or scallops. The

second is a feedrate modulation strategy that improves contact sta-
bility, optimizes cycle times, and more effectively removes workpiece
imperfections.

2.3.1. Spring rate modulation for improved burr removal

In deburring, large burrs, scallops, index tabs, and other imper-
fections should be removed from the part edge. Unfortunately, com-
plete removal of these imperfections is challenging. Even larger burrs
are small compared to the typical path accuracy of the manipulator, Fig. 5. Visualization of the feedrate modulation scheme from Eq. (22) with 𝑎𝑣 = 0.2,
𝑏𝑣 = 0.4 𝑓𝑑 .
and without knowledge of the workpiece geometry, a nominal force-
controlled deburring process cannot differentiate between a workpiece
imperfection and the workpiece itself. However, the SRAM method
offers two key advantages. First, the path error is reduced to negligible modulating the feedrate, more time can be spent on areas with greater
levels almost immediately once machining begins [15], enabling the need, and faster feedrates used otherwise. Second, feedrate modulation
detection and rejection of scallops and burrs. Second, the tuning pa- can enhance contact stability. There are situations where the tool loses
rameter 𝛾𝑎 provides a reliable estimate of the current path accuracy. If contact with the part or applies excessive force, despite the improved
the tool encounters an obstacle in the path and the path is estimated to performance of the SRAM method. If the feedrate is held constant,
be accurate, e.g., 𝛾𝑎 → 0, we can reasonably conclude that the obstacle sections of the workpiece may be unprocessed (in the case of contact
is a workpiece imperfection that should be removed. loss) or damaged (in the case of over-force). Modulating the feedrate
To achieve the desired logic, we propose a modification to the can slow the tool, giving the force controller more time to stabilize
admittance update equation from Eq. (3). While the admittance adap- on the workpiece edge and avoid potential contact-loss or over-force
tation term 𝛾𝑎 was previously used only to modulate virtual damping errors. Finally, with the spring constant modulation adjustments from
as 𝑏𝑎 ∕𝛾𝑎 , we propose additionally modulating the virtual stiffness as the previous section, the tool will avoid deflecting around large work-
[ ] piece imperfections such as burrs, scallops, or index tabs. Instead, it will
1 𝑏 𝑘𝑎
𝑎̈ = 𝑓𝑑 − 𝑓 − 𝑎 𝑎̇ − 𝑎 , (21) push through the imperfection to remove it entirely. Unfortunately, this
𝑚𝑎 𝛾𝑎 𝛾𝑎 [𝑗𝑘 1(−𝑎)] strategy, combined with a constant feedrate, causes the contact force to
where 𝑗𝑘 is a tunable shape parameter, and the Heaviside function increase drastically, resulting in excessive tool wear. Moreover, the tool
1(⋅) returns one when its input is positive and zero otherwise. Thus, will only spend a short time on the burr, which may be insufficient
the effective spring constant is 𝑘𝑎 when the admittance correction 𝑎 is to fully remove it. By adaptively slowing the feedrate, the force can
positive, and 𝑘𝑎 ∕𝛾𝑎 𝑗𝑘 otherwise. When the path estimate is poor, 𝑘𝑎 must be held more constant, and the tool will spend more time on the
be tuned to a small, negligible value. Otherwise, the force controller imperfection, removing it more effectively.
would be discouraged from leaving the nominal, low-accuracy path, The proposed feedrate modulation strategy depends on two main
and result in poor edge-tracking. With the scheme defined in Eq. (21), factors: the measured normal contact force error, 𝑓err = 𝑓 − 𝑓𝑑 , and
this behavior is retained: if the path estimate is poor, 𝛾𝑎 ≈ 1, and thus the accuracy of the path, as estimated by the admittance adaptation
𝑘𝑎 ∕𝛾𝑎 𝑗𝑘 ≈ 𝑘𝑎 . However, when the path estimate is good, we can penalize parameter, 𝛾𝑎 . The desired feedrate logic is as follows:
path deviations by setting 𝑘𝑎 much higher. By dividing by 𝑘𝑎 ∕𝛾𝑎 𝑗𝑘 as
1. If the desired force is too low, then the tool has lost contact with
𝛾 → 0, 𝑘𝑎 becomes large. The parameter 𝑗𝑘 allows for more precise
the part. In this case, the feedrate should be slowed to allow
adaptation tuning based on the application, relative to the damping
the force controller to find the workpiece again without leaving
modulation 𝑏𝑎 ∕𝛾𝑎 .
unprocessed sections of the edge.
Different behaviors are desired depending on whether the admit-
tance correction 𝑎 is positive or negative. If 𝑎 is negative, it indicates 2. If the desired force is too high, then:
that the tool has moved outwards from the nominal path, and has (a) If the path accuracy is poor, i.e., 𝛾𝑎 ≈ 1, then the tool
likely encountered a burr, scallop, or other imperfection that should is attempting to move into the part due to path error. In
be removed. Thus, a large spring constant is ideal to prevent the this case, the feedrate should be slowed to allow the force
tool from deflecting around the obstacle, and fully remove it instead. controller to move away from the part and stabilize the
However, if 𝑎 is positive, indicating that the tool has moved inwards contact.
from the nominal path, then the tool is following a pocket in the (b) If the path accuracy is good, i.e., 𝛾𝑎 ≈ 0, then the tool is
workpiece periphery. Although such a deviation of the part periphery in contact with an obstacle but will not move around it
is undesirable, the ideal outcome is nonetheless to ensure that the edge because of the logic from Section 2.3.1. In these cases, the
inside this pocket is deburred appropriately. Thus, the desired behavior feedrate must also be slowed to lower the contact force
is to have a low spring constant, allowing the tool to deflect to follow while fully removing the workpiece imperfection.
the edge, despite the deviation. Including the Heaviside term 1(−𝑎) in
the exponent of 𝛾𝑎 in Eq. (21) accomplishes this logic when 𝑎 is positive Thus, despite various situations leading to changes in 𝑓err , the solution
by resetting the effective spring constant to be 𝑘𝑎 ∕𝛾𝑎 0 = 𝑘𝑎 . always involves decreasing the feedrate when the force error magnitude
A downside of this proposed modification is that the contact force increases—regardless of 𝛾𝑎 .
may increase substantially when encountering obstacles. However, this This paper proposes the feedrate modulation curve shown in Fig. 5
issue is addressed by the second proposed improvement in the following to implement the desired logic. Mathematically, this curve is described
section. by the equation
𝑣nom
𝑣= (22)
2.3.2. Real-time feedrate modulation for improved contact stability | 𝑓err |1∕𝑎𝑣
1+| 𝑏 |
The second proposed improvement modulates the tool feedrate | 𝑣 |
based on process conditions, and offers many advantages for precision where 𝑣nom is the nominal feedrate, and 𝑎𝑣 and 𝑏𝑣 control the bell
robotic deburring applications. First, modulating the feedrate can im- curve shape. The parameter 𝑎𝑣 controls the steepness of the transition
prove cycle times. With constant feedrate, a conservative speed must distance, with smaller values giving a steeper transition, and 𝑏𝑣 is the
be selected, dictated by the largest burrs or scallops on the part. By distance to 𝑣 = 12 𝑣nom , controlling the width of the bell curve.

6
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

The two proposed changes, spring constant and feedrate modula-

tion, enable the SRAM-D method to perform robotic deburring with
greater contact stability than the original framework, while simultane-
ously being more effective at removing peripheral imperfections such as
large burrs, scallops, or index tabs. The following sections will demon-
strate how the method can be implemented on an articulated robot and
provide details on calibrating and tuning the various parameters of the
SRAM-D method.

3. Application of method

The proposed method is implemented on a robotic machining sys-

tem to validate the SRAM-D improvements. This section describes the
method implementation, with Section 3.1 describing the experimental
platform used for validation, Section 3.2 demonstrating a calibration
procedure for the SRAM tool models, and Section 3.3 providing detailed
guidelines to set the tuning parameters of the SRAM-D framework.

3.1. Experimental platform

Experimental validation is performed on a 6-DOF Denso VS-6556 W

serial industrial robot, as depicted in Fig. 6. The workcell and robot
end-effector have been designed for robotic machining experiments and
feature:

• A 350 W NSK E3000 NR-3060S precision spindle;

• A vacuum clamp table, for rapid clamping of planar sheet-metal
workpieces;
• A Diversitech downdraft table, to safely remove dust and swarf;
• An ATI Gamma IP60 six-axis force/torque sensor, to measure Fig. 6. Deburring workcell used in the validation experiments.
end-effector forces;
• A Micro-epsilon scanCONTROL 2950-50 laser scanner, for work-
piece registration; provided in [53]. This method is less accurate, with a mean position
• A stereo camera system consisting of two Basler error of 2.5 mm (0.098’’), but can be completed much more quickly, in
a2A3840-13gcPRO cameras, for faster but less accurate workpiece approximately 15 s.
registration. Deburring tests are conducted on planar aluminum sheet metal
workpieces. Two sample workpieces are defined, as shown in Fig. 7.
The robotic manipulator has been kinematically calibrated and
Both are created through a nibbling process from 6061-T6 aluminum
achieves approximately 0.25 mm absolute accuracy within the re-
gion of the vacuum clamp [19]. To compute the inverse kinematics sheet metal with a thickness of 1.8 mm (0.071’’). The nibbling process
of the calibrated kinematic model, which lacks the joint alignment leaves behind a rough edge, a tear burr on the lower side, and scal-
typically used in closed-form solutions, we employ the fast and ro- loping between die punches. The first workpiece (Fig. 7a) is a small
bust iterative QuIK inverse kinematics method [51]. The robot is rectangular coupon with simple geometry, and is used to calibrate the
controlled using PID controllers with added gravity and friction com- tool models in the following section. The second workpiece (Fig. 7b)
pensation as described in [42,52]. Real-time control is performed from has a more complex shape that includes sharp and rounded internal and
Matlab/Simulink 2022b, using the Quarc hardware-in-the-loop system external corners to assess deburring performance in various scenarios.
developed by Quanser. The deburring tool is a 3/4’’ 120-grit flapwheel This workpiece also features three deliberate scallops, with a height and
tool, rotating at 22 000 rpm. The path planning process from [53] is width of 0.32 mm (0.013’’), 0.64 mm (0.025’’), and 1.27 mm (0.050’’),
used to program the contouring path. The path commands a desired respectively, which will be used to evaluate the deburring controllers’
workpiece-tool contact force of 𝑓𝑑 = 2.5 N; however, this force is ability to remove workpiece imperfections.
reduced to 1.25 N around corners to prevent rounding. Data collection
and controller updates are processed at 1000 Hz, however the SRAM- 3.2. Tool model calibration
D adaptation block is run more slowly at 25 Hz to limit the size of
the accumulated data. Aliasing is avoided during downsampling and The tool models discussed in Section 2.2.2 and Section 2.2.3 must be
upsampling using 2nd-order Butterworth lowpass filters at 6 Hz. More calibrated to leverage the SRAM framework effectively. For the contact
complete implementational details for this system are available in [15]. model, the tool stiffness 𝑘𝑡 is required, and for the wear model, the wear
Two modalities are available for workpiece registration. The laser- rate 𝜓𝑡 , shape parameter 𝑗𝑡 , and the lowpass cutoff frequency 𝐹𝑡 are re-
scanner system, calibrated to the robot’s kinematic model, can localize quired. The following section provides a straightforward methodology
workpieces by scanning over the part and building a 3D point cloud to calibrate the tool model and accurately determine these parameters.
to fit the workpiece geometry. This method is very accurate, giving a
mean position error of 0.4 mm (0.015’’). However, the scanning and 3.2.1. Tool stiffness calibration
fitting process is lengthy, requiring approximately 63 s to complete. The tool stiffness is a lumped parameter that incorporates both the
The second method uses the camera system for localization. The part is flexibility of the tool and of the robot itself. It may not be possible to
photographed, and the image undergoes a multi-stage process including approximate this parameter accurately—the tools themselves can differ
rectification, lens warp correction, background subtraction, and edge greatly from their manufacturing, and the robot’s stiffness changes
detection, to isolate the workpiece contour. This contour is then fitted based on pose and orientation. Fortunately, the tool stiffness is also not
to workpiece geometry using an ICP algorithm. Further details are critically important to the SRAM algorithm. Since the applied force 𝑓 is

7
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Fig. 8. Visualization of the tool stiffness testing path. Phase of the force oscillations
is randomized for each test.

Fig. 9. Observed and predicted spring forces on the validation data.

as the vector 𝑟𝑖 . Thus, these residuals are uncorrelated to time and only
relate to the applied force 𝑓𝑖 according to the spring rate
Fig. 7. Sample calibration and experimental validation workpieces. 𝑓 𝑖 = 𝑘𝑡 𝑟𝑖 . (24)

Eight workpieces are processed in this manner. Six are used as

training data, while the remaining two are used for validation. Data
approximately constant, a change in the stiffness 𝑘𝑡 can be interpreted
is collected at 1000 Hz. High-frequency data is removed from 𝑓𝑖 and 𝑟𝑖
equivalently as a change in the initial tool radius 𝑟𝑡 , since both act
by filtering with a 4th-order Butterworth filter with a cutoff frequency
collinearly in the estimation of 𝐩̃ 𝑐𝑖 in Eq. (5). Thus, small errors in 𝑘𝑡
̄ at 6 Hz, and downsampled to 25 Hz. Samples with 𝑓𝑖 < 𝑓thresh where
can be handled without significantly reducing path accuracy.
𝑓thresh = 0.5 N are ignored. The residuals 𝑟𝑖 and forces 𝑓𝑖 are collected
Tool stiffness depends on several factors, including tool type, robot
for all segments and solved using least-squares regression. The spring
pose, tool speed, and workpiece thickness. As such, an offline iden-
constant is identified to be 7.61 N/mm, as listed in Table 1. Fig. 9 shows
tification is required if any of these items change significantly. In
the fitted data and residuals. The resulting fit has a validation 𝑅2 value
this paper, the tool radius and workpiece thickness are constant, and
of 0.83, capturing most, but not all variance in the data. This slightly
deburring occurs within a smaller subset of the robot workspace. Thus,
imperfect fit is expected due to factors such as pose-dependent robot
tool stiffness may be considered approximately constant.
stiffness and tool-to-tool variance. However, the accuracy is sufficient
Identification of the tool stiffness is performed using the rectangular
for the application. More complex models could seek to change the tool
workpiece from Fig. 7a, mounted to the vacuum clamp. A deburring
radius in real-time based on the robot pose, but this approach would
path is planned around the part, using an admittance force controller
introduce significant additional complexity and is beyond the scope of
operating along the normal axis 𝐮𝑛 , as described previously. Feedrate is
̄ the current work.
constant at 5 mm/s. The desired force is then modulated sinusoidally
as shown in Fig. 8 between 0 and 𝑓max , as 3.2.2. Tool wear calibration
( ) Tool wear calibration identifies the wear rate 𝜓𝑡 , the shape parame-
𝑓𝑑 (𝑡) = 𝑓max ∕2 + sin 2𝜋𝜔𝑡 + 𝜙 , (23)
ter 𝑗𝑡 , and the wear frequency 𝐹𝑡 . As with the tool stiffness, these wear
where 𝑓max = 5 N, twice the nominal tool force, 𝜔 is set to 0.25 rad/s— parameters will depend on tool type, tool speed, and part thickness.
fast enough to allow for multiple cycles per edge, but slow enough that Identification should be performed if any of these factors are changed.
the force controller can accurately track the setpoint. Finally, the phase A rectangular coupon is placed on the vacuum clamp, and a de-
𝜙 is set randomly for each experiment. burring path is planned around the part using a simple admittance
As the tool contours the part under admittance control, the admit- force controller with a feedrate of 5 mm/s. The applied force is varied
tance correction 𝑎 is logged. The data is then split into parts corre- throughout the path by selecting a random integer value between 1 N
sponding to each workpiece edge. For each of the four linear segments, and 8 N and holding this force for 4 s, then repeating until the end
the data is first detrended to remove any linear trends with respect to of the contouring path. Contact forces higher than 8 N were avoided,
time in the data, which are introduced from slight rotations in the part as they caused extreme wear and stalled the spindle. This procedure
registration or from tool wear. The detrending can be accomplished by results in force profiles as shown in Fig. 11, which allow the tool
fitting a first-order line to the admittance correction 𝑎 and time 𝑡, then wear to be observable at different contact forces. Eight unique tests are
removing this trend from the data to leave only the residuals, denoted performed in this manner. Additionally, we perform seven tests holding

8
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Table 1
Identified tool properties.
Value RMSEa 𝑅2 a
Spring constant 𝑘𝑡 7.61 N∕mm 0.59 N 0.83
Wear rate 𝜓𝑡 0.94 μm∕Njt s
Wear exponent 𝑗𝑡 2.01 0.14 mm 0.75
Wear frequency 𝐹𝑡 239 Hz
a RMSE and 𝑅2 are reported on the validation datasets.

Fig. 11. Observed normal residuals and wear in validation.

Fig. 10. Effect of the wear lowpass filter cutoff frequency, 𝐹𝑡 , on the accuracy of the
wear model.

a constant force for the entire run, with the force set respectively at
each integer value from 1 N to 7 N. The seven constant force tests and
five of the random force tests were used for training data, while the
remaining three random force tests were set aside as a validation set.
The tool parameters are identified with the acquired data through Fig. 12. Normal SRAM residuals with and without wear model.
a nonlinear least-squares optimization using the SRAM equations, as in
Section 2.2. For a given set of tool parameters 𝜓𝑡 and 𝑗𝑡 on a given
test, the SRAM update equations above are used, with the force and 3.3. Parameter tuning
position data from the test, to iteratively converge to the true workpiece
location and tool diameter, while holding the tool wear parameters 𝜓𝑡 To leverage the SRAM-D method, the tuning parameters must be se-
and 𝑗𝑡 constant. This process is repeated for each test and the residuals lected for optimal performance in the application. Typically, parameter
𝐞𝑖 are then collected into a single residual vector which is fed to a tuning must be done experimentally, and is time consuming and rep-
trust-region nonlinear least-squares solver, implemented using Matlab’s resents a barrier to industrial adoption. However, many parameters in
lsqnonlin function, and run until convergence. SRAM and SRAM-D are motivated by physically measurable quantities,
The measured forces 𝑓𝑖 must be lowpass filtered to remove forces allowing an initial estimate to be readily obtained. Other parameters
that do not represent actual physical workpiece-tool interactions. This affect algorithm performance only indirectly, and thus can be tuned
filter is implemented as a 2nd-order Butterworth filter with a cutoff offline, by performing a single deburring run, then ‘‘replaying’’ the con-
frequency 𝐹𝑡 . The optimization process above is repeated at logarithmi- vergence of algorithm in simulation. In this way, the parameter tuning
cally increasing values to determine the most appropriate value for 𝐹𝑡 . for the SRAM algorithm can be performed with minimal experimental
For each frequency, a scalar cost is computed as the sum-of-weighted- testing.
squared-residuals across all tests. This data is plotted in Fig. 10, and This section provides guidelines to aid the reader in optimally
shows that from 𝐹𝑡 ≈ 30–240 Hz, the cost is approximately constant, tuning the SRAM-D framework for robotic deburring. The guidelines
with the minimum cost at 𝐹𝑡 = 239 Hz. This result is consistent with are broken into six sections below, covering the tuning of the contact
model, algorithm convergence, outlier detection, admittance force con-
our expectations, as the natural frequencies of the robot are typically
troller and adaptive modulation, spring constant modulation, and fee-
observed in the range of 10–26 Hz and cause the most significant
drate modulation. The selected parameters are summarized in Table 2.
physical workpiece-tool interactions. Meanwhile, the other large exci-
tation in the force signal is at the spindle frequency of 22 000 rpm,
3.3.1. Contact model
or 366 Hz. Force frequency components in this range are imparted
The contact model requires two parameters: the tool stiffness 𝑘𝑡 ,
from the spinning of an eccentric tool, but do not represent real
calibrated in the previous section, and the in-contact force threshold,
workpiece-tool interaction forces and are uncorrelated to tool wear.
𝑓thresh , which is used to detect when the tool contacts the part. The
The resulting model parameters are summarized in Table 1. Fig. 11 parameter 𝑓thresh should be set as low as possible, while remaining
shows the normal component of the SRAM residuals without wear higher than the accuracy of the force–torque sensor. The load cell is
compensation, computed using the identified the fully converged state accurate within ±0.2 N in the current setup, and thus 𝑓thresh is set to
𝛒, compared to the predicted wear rates. Additionally, Fig. 12 plots the 0.5 N.
distribution of the normal component of the model residuals 𝐞𝑖 , before
̄
and after the wear model is applied. Overall, the model captures the 3.3.2. Algorithm convergence
most significant effects of tool wear, and significantly reduces the mean Three parameters influence the SRAM algorithm convergence: the
path error in the SRAM models from 0.42 mm to 0.14 mm in validation, initial state covariance 𝐏0 , the residual covariance 𝐒𝑖 , and the update
giving a fit of 𝑅2 = 0.74. The remaining errors can be attributed damping factor 𝛌. The first two parameters are measurable variances
to limitations in robot accuracy, workpiece errors, and the contact readily estimated from experimental data. However, it may be desirable
model, which does not capture spring nonlinearity. Nonetheless, the to tweak the values to achieve a given behavior—reducing the variance
model reduces the SRAM residual error to within the region of the of 𝐏0 leads the algorithm to rely more on the real-time observations,
robot accuracy itself, beyond which further improvements may not be while reducing the variance of 𝐒𝑖 instead gives more weight to the
feasible. initial state estimate.

9
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Table 2
Selected tuning parameters for SRAM-D controller.
In-contact force threshold, 𝑓thresh 0.5 N
[ ]
Initial registration covariance, 𝐏0 diag 32 32 32 22 0.00252 a
[ 2 2
] 2
Residual covariance, 𝐒𝑖 diag 0.1 0.1 mm
[ ]
Update damping factor, 𝛌 diag 0.1 0.1 30 105 105 a
Distance weight transition width, 𝑎𝑑 1 std
Distance weight transition center, 𝑏𝑑 3 std
Time weight transition width, 𝑎𝑡 1 s
Time weight transition center, 𝑏𝑡 1.5 s
Virtual admittance mass, 𝑚𝑎 2.5 kg
Virtual admittance damping, 𝑏𝑎 0.225 Ns/mm
Virtual admittance spring, 𝑘𝑎 0.05 N/mm
Initial admittance std., 𝜎𝑎0 3 mm
Minimum admittance std., 𝜎𝑎,min 0.01 mm Fig. 13. Visualization of the parameters controlling the admittance tuning factor, 𝛾𝑎 .
Admittance tuning shape factor, 𝑗𝑎 0.5
Minimum admittance tuning, 𝛾𝑎,min 0.05
Spring modulation shape factor, 𝑗𝑘 1.5
Nominal feedrate, 𝑣nom 6 mm/s machining behavior significantly and should be tuned experimen-
Feedrate transition width, 𝑎𝑣 0.2 tally based on the application and tooling. For our manipulator
Feedrate transition center, 𝑏𝑣 0.4 𝑓𝑑 and tooling, 𝑏𝑎 = 0.225 Ns/mm was selected experimentally
a 𝐏 has units of mm2 for 𝛿𝑥, 𝛿𝑦, and 𝛿𝑟, deg2 for 𝜃 and (μm∕Njt s)2 for 𝛿𝜓 . 𝛌 as the lowest damping possible before tool chatter increased to
0 𝑡
uses the reciprocal of these same units (mm−2 , deg−2 and (μm∕Njt s)−2 . unacceptable levels.
• The nominal spring 𝑘𝑎 should be set to a small positive value,
large enough to draw the tool to the nominal path in the absence
of other factors, without significantly affecting the contact forces
The damping parameter 𝛌 should be tuned to give reasonable con-
at expected path errors. For instance, nominal path errors of
vergence speeds. Initially, we recommend using 𝛌 ≈ 𝐏−1 0
∕𝐹𝑠 , where
approximately 3 mm may be expected in our case, and the desired
𝐹𝑠 is the SRAM update rate, which approximates the correct order of
contact force is 𝑓𝑑 = 2.5 N. Therefore, 𝑘𝑎 = 0.05 N/mm is selected,
magnitude, after which simulated tuning can fine-tune the value. The
which induces 0.15 N at typical path errors—small enough not to
damping 𝛌 can also be used to prioritize the adaptation of specific states
affect machining quality, but strong enough to draw the tool to
over others. For example, when the tool initially contacts a workpiece, the nominal path without other interactions.
there is an observability null-space between the workpiece translation
and the tool radius. From the algorithm’s perspective, the part could be The remaining parameters 𝜎𝑎0 , 𝜎𝑎,min , 𝑗𝑎 , and 𝛾𝑎,min control the SRAM-D
shifted, or the tool radius size could change. Preferentially converging adaptive tuning. These tuning values are visualized in Fig. 13 and can
𝛿𝑥 and 𝛿𝑦 over 𝛿𝑟, by increasing the value of the 4th element of 𝛌, be adjusted as follows:
generally produced better overall path accuracy. The values used in
• The initial admittance standard error 𝜎𝑎0 represents the uncer-
the current experiment are listed in Table 2 and were estimated from
tainty of the admittance correction 𝑎 with the initial state co-
experimental data before being fine-tuned in simulation.
variance, 𝐏0 . It can be estimated from the average path error
of a nominally planned contouring path, given your registration
3.3.3. Outlier detection method. Alternatively, it can be estimated by substituting the
Outlier detection relies on two weighting schemes: distance-based initial state covariance 𝐏0 into Eq. (19). We use 𝜎𝑎0 = 3 mm, based
and time-based, each with two parameters controlling the steepness on typical nominal path error.
(𝑎𝑑 and 𝑎𝑡 ) and center (𝑏𝑑 and 𝑏𝑡 ) of the transition. For the distance- • The minimum admittance standard error 𝜎𝑎,min represents the
based weighting, the aim is to include ‘‘reasonable’’ points and exclude lowest uncertainty possible for the admittance correction 𝑎. It
statistically improbable ones. We propose 𝑎𝑑 = 1 std and 𝑏𝑑 = 3 std, can be estimated from the manipulator’s repeatability or the
which includes points within 2.5 std of the mean (approximately 99% workpiece profile accuracy—whichever is higher. In this paper,
of the data, if normally distributed) and unweights points further from 𝜎𝑎,min is set near the robot repeatability at 0.01 mm.
the norm. The time-based weighting aims to prevent novel points from • The shape factor 𝑗𝑎 controls the interpolation of 𝛾𝑎 . We recom-
being excluded until sufficient ‘‘new’’ points have been gathered to mend a starting value of 𝑗𝑎 = 0.5. Changes in 𝑗𝑎 are visualized
apply the distance weighting accurately. In our testing, 𝑎𝑑 = 1 s and in Fig. 13: increasing 𝑗𝑎 results in higher 𝛾𝑎 for intermediate 𝜎𝑎 ,
𝑏𝑑 = 1.5 s yielded good results. We do not expect these outlier detection while decreasing it leads to lower intermediate 𝛾𝑎 .
parameters to require significant modification in other applications. • The minimum admittance tuning 𝛾𝑎,min establishes an absolute
lower bound on 𝛾𝑎 , to ensure that the admittance tuning remains
3.3.4. Admittance force controller tuning within reasonable and stable limits. Even with perfect registra-
Several parameters control the force controller’s tuning. The virtual tion, a certain amount of force control is required to maintain
mass-damper-spring parameters of the admittance controller determine stable workpiece-tool contact. If this minimal force control has an
nominal behavior and can be tuned as follows: associated damping 𝑏𝑎,max , then an appropriate tuning is 𝛾𝑎,min =
𝑏𝑎 ∕𝑏𝑎,max . We use 𝛾𝑎,min = 0.05, resulting in a maximum damping
• The virtual mass, 𝑚𝑎 , should be set to limit tool accelerations of 𝑏𝑎,max = 4.5 Ns/mm.
to reasonable values. A good starting value is the true end-
effector mass, as setting the virtual mass much smaller can pro- 3.3.5. Spring constant modulation tuning
duce accelerations that the manipulator cannot deliver. On our The spring constant modulation scheme in Section 2.3.1 is parame-
manipulator, 𝑚𝑎 = 2.5 kg provided good performance without terized by the shape factor 𝑗𝑘 that allows the modulation of the virtual
excessive accelerations. spring constant to be separated from the virtual damper tuning. Tuning
• The nominal damping should be the lowest value that still results of 𝑗𝑘 is easiest accomplished by considering the ideal admittance tuning
in a stable workpiece-tool contact, as the damping can only at full convergence, i.e., when 𝛾𝑎 = 𝛾𝑎,min . From this maximum spring
increase as the controller tuning 𝛾𝑎 is adjusted throughout the run. constant 𝑘𝑎,max , the parameter 𝑗𝑘 can be determined as
Lower values improve edge-tracking, while higher values increase 𝑘𝑎 log(𝑘𝑎 ∕𝑘𝑎,max )
workpiece-tool contact stability. The controller damping impacts 𝑘𝑎,max = ⟹ 𝑗𝑘 = . (25)
𝛾𝑎,min 𝑗𝑘 log(𝛾𝑎,min )

10
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

In this paper, we set 𝑗𝑘 = 1.5, which corresponds to 𝑘𝑎,max = 4.5 N/mm. 2. SRAM: The original SRAM algorithm, as introduced by Lloyd
Increasing the stiffness further destabilizes the force controller but does et al. in [15], and tuned according to Table 2.
not significantly increase the perceived workpiece-tool stiffness, as the 3. ADMIT: A nominal admittance controller without SRAM up-
tool stiffness 𝑘𝑡 is already of similar magnitude. dates, spring modulation, or feedrate modulation. The controller
is tuned with 𝑚𝑎 = 2.5 kg, 𝑏𝑎 = 0.75 Ns/mm, and 𝑘𝑎 =
3.3.6. Feedrate modulation tuning 0.05 N/mm, which represents the best trade-off of edge-following
The final tuning parameters 𝑣nom , 𝑎𝑣 , and 𝑏𝑣 adjust the feedrate performance and contact stability in pre-testing, without the
modulation curve shown in Fig. 5. The nominal feedrate 𝑣nom must improvements offered by SRAM.
be tuned to the application. If set too fast, the part may not be fully 4. Manual: The workpiece is also deburred by a human worker for
deburred, but the cycle times may be unnecessarily prolonged if set comparison.
too slowly. For the current experiments, we found that 𝑣nom = 6 mm/s
The experimental setup described in Section 3.1 is highly optimized
resulted in a good balance of efficiency and quality. for accuracy, featuring a kinematically calibrated robot and laser scan-
The remaining parameters, 𝑎𝑣 and 𝑏𝑣 , control the transition of the ner. The manipulator’s smaller size also provides higher repeatability
feedrate curve, with 𝑎𝑣 determining the transition steepness and 𝑏𝑣 than most industrial manipulators. Therefore, localization accuracy is
controlling the transition center, as in Fig. 5. An ideal curve allows for higher with this system than a typical industrial workcell. To evaluate
‘‘normal’’ force deviations without slowing the feedrate, but reduces the performance of SRAM in a variety of potential implementations,
the speeds for larger force errors caused by contact loss or large forces. each controller is tested using one of the four registration profiles
Additionally, the feedrate transition cannot be too steep, or changes identified in Table 3. The tool radius is accurately known beforehand,
in the feedrate may occur faster than the robot can respond, resulting as all tests are done with a new tool surface to avoid inconsistent
in low-frequency feedrate oscillations. In the current work, the curve results. An artificial perturbation to 𝑟𝑡 is added in all profiles, sampled
𝑎𝑣 = 0.2, 𝑏𝑣 = 0.4𝑓𝑑 provided a good trade-off of these constraints. from a normal distribution with a standard deviation of 1 mm to
simulate an unknown tool radius typical of actual operations.
4. Results and discussion The LASER registration profile uses the robot-mounted laser system
to localize the part, and after the tool radius perturbation, results in a
The proposed SRAM-D method is evaluated in a comprehensive typical mean path error of 0.7 mm. This profile is representative of any
robotic deburring experiment against two nominal controllers and un- similar workcell with an accurate, small, calibrated robot using laser lo-
der four registration profiles with varying workpiece registration accu- calization. The CAM registration profile utilizes the camera system from
racy. Section 4.1 outlines the testing methodology. Section 4.2 demon- Section 3.1, and with the radius perturbation, has a typical mean path
strates the convergence properties of the SRAM-D method and quanti- error of 1.5 mm. This profile is representative of any similar workcell
fies the path accuracy improvements provided. Section 4.3 investigates with an accurate, small robot using camera localization. Additionally,
the contact stability of the tested algorithms, and Section 4.4 measures we define the CAM+ and CAM++ profiles, which use the camera system
their effectiveness in removing scallops, burrs, and other workpiece as in CAM, but introduce additional synthetic perturbations to 𝛿𝑥, 𝛿𝑦,
imperfections from the part edge. Section 4.5 compares tool wear and 𝜃 to artificially increase the error. CAM+ adds a small perturbation,
for the tested methods, and Section 4.6 evaluates cycle times and resulting in a mean path error of 3.4 mm, simulating a workcell with
corresponding workcell efficiency for the tested registration methods. camera localization and either a smaller manipulator without kinematic
calibration, or a larger manipulator with kinematic calibration. CAM++
4.1. Methodology introduces a more significant perturbation, resulting in a mean path
error of 5.8 mm and simulating a workcell with camera localization and
Deburring tests are performed with the workpiece in Fig. 7b. Four a larger, uncalibrated robot, or other imprecise registration method.
deburring methodologies are evaluated: Each combination of controller and registration profile is evaluated
in five tests in different locations on the vacuum clamp shown in Fig. 6,
1. SRAM-D: The deburring SRAM algorithm, as proposed in the averaging the results for greater statistical significance and resulting in
current paper, with added spring constant and feedrate modu- a total of 60 robotic deburring tests. The manually deburred samples
lation modules, and tuned according to Table 2. are also averaged over five tests. The random perturbations of the

Table 3
Summary of testing registration profiles and associated accuracy.
Perturbation std. Measured std. Path error
𝛿𝑥 𝛿𝑦 𝜃 𝛿𝑟 𝛿𝑥 𝛿𝑦 𝜃 𝛿𝑟 ave. max
Description Sample applications [mm] [mm] [deg] [mm] [mm] [mm] [deg] [mm] [mm] [mm]
LASER Accurate but slow registration ∙ Laser localization with small 0 0 0 1 0.17 0.20 0.14 0.83 0.67 1.8
with costly laser scanner. calibrated robot.

CAM Fast but less precise registration ∙ Camera localization with small 0 0 0 1 1.9 1.1 0.69 0.82 1.5 5.2
with inexpensive camera system. calibrated robot.

CAM+ Fast but less precise registration ∙ Camera localization with small 2.5 2.5 1.25 1 2.9 1.5 1.8 0.51 3.4 9.5
with inexpensive camera system uncalibrated robot.
with small added error. ∙ Camera localization with large
calibrated robot.

CAM++ Fast but less precise registration ∙ Camera localization with large 5 5 2.5 1 3.9 3.5 2.8 0.76 5.8 15.1
with inexpensive camera system uncalibrated robot.
with larger added error. ∙ Imprecise registration: rough jigging,
manual placement, etc.

Measured standard error and path error use the converged SRAM state as the ground truth. Reported standard deviations are the principle deviations of a multivariate normal
distribution with assumed zero mean. Perturbed and measured standard deviations differ due to small sample sizes.

11
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Fig. 14. Visualization of state convergence during sample SRAM-D deburring, with CAM+ registration profile.

registration profiles are held constant between the three controllers to

ensure a consistent comparison. In the following sections, we present
and discuss the results of this testing.

4.2. Algorithm convergence & path error

As the tool contours the part during deburring with the SRAM and
SRAM-D controllers, the part registration and tool radius are updated
in real-time to reduce path error. Simultaneously, the admittance con-
troller is adapted according to the live covariance estimate. This section
examines the algorithm convergence and demonstrates how SRAM and
SRAM-D improve path accuracy.
Consider Fig. 14, which illustrates the SRAM registration estimate
and deburring path from a sample test using the CAM+ registration pro-
file and the SRAM-D controller. Initially, in Fig. 14a, the estimated part
location (dashed purple) is inaccurate. Accordingly, the force controller
adaptation 𝛾𝑎 is close to one for low damping and optimal edge-finding
ability. However, when the tool contacts the part and begins moving
along the first edge, as shown in Fig. 14b, the part registration is
adjusted to fit the observed workpiece-tool contacts. At the same time,
the force controller rapidly transitions to 𝛾𝑎 → 0, increasing damping
for more stable workpiece-tool contact. Note that the transition to a Fig. 15. Path accuracy of sample deburring runs from each controller and registration
damped tuning occurs along the first edge, before the controller has profile combination.
acquired any observations that give information as to the horizontal
part position or the tool radius. The state covariance is projected onto
the current admittance axis, allowing the controller tuning to occur Fig. 15 plots the path error of sample tests for each combination of
based on the local variance in the admittance axis itself. Thus, the controller (SRAM-D, SRAM, and ADMIT) and every registration profile
controller can tune for high damping when it is confident in the path (LASER, CAM, CAM+, and CAM++). The figure shows the path error
along a given axis, even if the total state has not converged or if the of the SRAM and SRAM-D controllers reduced to sub-millimeter levels
registration along another axis is poor. As the tool turns the first corner, almost immediately after the run starts and the tool makes contact
the controller reverts to low damping for better edge tracking while with the part, regardless of initial registration accuracy. Conversely, the
converging the part’s horizontal offset. After this corner, in Fig. 14c, the ADMIT controller path error remains uncorrected, and the controller
remaining registration state variables 𝛿𝑥, 𝜃, and 𝛿𝑟 are fully converged. must use force control to compensate for the large, changing path
For the remainder of the deburring run, the controller operates with errors. For the LASER registration profile, the improvement in path
low 𝛾𝑎 and high damping, ensuring good contact stability. Therefore, accuracy is minor, but for the CAM, CAM+, and CAM++ profiles, the
throughout the entire process, the admittance controller is optimally improvement is more pronounced. In the CAM++ profile, the ADMIT
tuned: when registration is poor and edge-finding capability is required, controller was emergency stopped partway through the test due to
damping is automatically lowered, and otherwise it is increased for unexpected workpiece-tool collisions and high forces from the excessive
better contact stability and machining quality. path error. Meanwhile, the SRAM and SRAM-D controllers complete the
As new information is acquired, the commanded path (shown in run with negligible error.
dashed red) is updated from the initial path (shown in dashed white) The averaged path error across all 60 deburring tests is summarized
based on the updated state information. Because these corrections are in Table 4. This table shows that the SRAM and SRAM-D controllers are
applied in real-time, the commanded path never deviates significantly able to effectively reduce the root-mean-squared error (RMSE) of the
from the ideal path. Thus, path error is minimal for the run—even be- path to negligible levels across all registration profiles. In the CAM+
fore the state fully converges. This minimal error is evident in Fig. 14c, and CAM++ profiles, path error does increase slightly compared to
as the commanded path (dashed red) never deviates significantly from the LASER profile. However, this increase is due to the larger initial
the force-controlled path (thick color-coded line), despite substantial error when the tool initially finds the first workpiece edge, and is
error in the initial path (dashed white). unavoidable as registration accuracy decreases. Otherwise, the SRAM-D

12
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Table 4
Averaged path accuracy and contact stability errors.
Path errora [mm] Contact-loss [mm] Over-force [mm] Chatter power [N2 ]
SRAM-D 0.37 (↓ 45%) 0.06 (↑ ∞%) 0.21 (↑ 121%) 0.18 (↓ 31%)
LASER SRAM 0.41 (↓ 41%) 0.11 (↑ ∞%) 0.05 (↓ 49%) 0.27 (↑ 4%)
ADMIT 0.65 0.00 0.09 0.26
SRAM-D 0.31 (↓ 80%) 0.14 (↓ 97%) 0.10 (↓ 98%) 0.08 (↓ 74%)
CAM SRAM 0.48 (↓ 66%) 0.54 (↓ 87%) 0.15 (↓ 97%) 0.21 (↓ 35%)
ADMIT 1.64 4.12 4.84 0.32
SRAM-D 0.53 (↓ 84%) 0.11 (↓ 99%) 0.11 (↓ 100%) 0.12 (↓ 51%)
CAM+ SRAM 0.48 (↓ 86%) 0.35 (↓ 98%) 0.00 (↓ 100%) 0.09 (↓ 65%)
ADMIT 3.20 16.1 28.3 0.25
SRAM-D 0.63 (↓ 88%) 1.44 (—) 0.12 (—) 0.19 (—)
CAM++ SRAM 0.66 (↓ 87%) 3.28 (—) 0.06 (—) 0.13 (—)
ADMITb 5.8b (FAIL) (FAIL) (FAIL)
SRAM-D 0.4 (↓ 84%) 0.10 (↓ 98%) 0.14 (↓ 99%) 0.13 (↓ 53%)
Averagec SRAM 0.51 (↓ 82%) 0.34 (↓ 95%) 0.07 (↓ 99%) 0.19 (↓ 32%)
ADMIT 2.82 6.75 11.1 0.28

Percentage given as relative error to the nominal path (for path error), and to the nominal admittance controller ADMIT (force contact-loss, over-force, and chatter power).
a
Path error is reported as the average of the root-mean-squared error (RMSE) across the 5 sample positions.
b
The nominal admittance controller ADMIT failed during the CAM++ registration due to excessive workpiece-tool forces. However, the path error is still known since it does not
change during the run.
c
Average improvement over the nominal ADMIT controller. Averages omit the CAM++ profile which failed for the ADMIT controller, and thus underestimates improvements.

controller lowered the path error by 45% compared to ADMIT in the In the current setup, the manipulator natural frequencies range
LASER registration profile, 80% for CAM, 84% for CAM+, and 88% for from approximately 10 to 26 Hz. Therefore, we define the contact-
CAM++. The nominal SRAM controller exhibited similar performance. loss and over-force metrics based on a force signal lowpass-filtered at
This reduced path error allows the controllers to achieve better contact 8 Hz, while chatter power is defined as the signal power in the 8–
stability in deburring, and use a more damped controller tuning that 30 Hz range. Filtering is performed with 2nd-order Butterworth low-
improves finish quality and more effectively removes scallops and and high-pass filters, applied non-causally to prevent phase shift using
burrs. The improvements in contact stability are discussed further in the Matlab’s filtfilt function. Additionally, contact-loss and over-
the following section. force are not considered near the artificial scallops on the workpiece’s
left side, as their size exceeds what is typically encountered in normal
4.3. Contact stability applications, and forces vary more significantly as a result without
necessarily representing abnormal behavior.
The stability of the workpiece-tool contact force significantly im- Fig. 16 shows the measured and commanded contact forces, 𝑓 and
pacts the final machined part quality. Excessive force can damage the 𝑓𝑑 , during a sample of 12 of the 60 deburring tests performed. The
workpiece, compromise its peripheral accuracy, and increase tool wear, low-frequency trends of the measured force, lowpass-filtered at 8 Hz,
while insufficient force does not deburr the edge properly. Additionally, are shown in dark blue, while the light-blue overlay is lowpass-filtered
force oscillations at improper frequencies can degrade surface finish at 30 Hz to show higher-frequency signal elements and visualize chatter
and cause premature tool wear. power. Contact-loss error is highlighted in purple, and over-force error
However, summary statistics such as root-mean-squared error is highlighted in orange. The plot visualizes the improvements in con-
(RMSE) or mean absolute error (MAE) of force error do not correlate tact stability provided by the SRAM framework. In the CAM and CAM+
well with workpiece quality. Smaller force errors are expected when profiles, the ADMIT contact forces fluctuate significantly as the force
the tool encounters burrs or scallops, causing normal force deviations. controller struggles to accurately track the large path errors without
Issues only arise with large deviations in the lower frequencies of the path corrections and adaptive tuning of the SRAM methods. In
the force error signal. In the higher frequency bands of force error, the CAM++ profile, the ADMIT test failed due to excessive force. The
larger signal power near the structural frequencies of the manipulator ADMIT controller avoided contact-loss and over-force error only in the
correlate to chatter, tool wear, and reduced part quality. Thus, to LASER profile, thanks to the highly accurate, sub-millimeter accuracy
more accurately capture the relationship between contact force and registration.
workpiece quality, and summarize the test results more meaningfully, Conversely, across all registration profiles, contact-loss and over-
we utilize the following metrics: force error is almost entirely eliminated from the path in the SRAM
and SRAM-D controllers. The only remaining contact error in these
1. Contact-loss: Contact-loss error describes when the tool is not controllers is momentary contact-loss error at the beginning of the path,
deburring the workpiece edge effectively, and occurs when the as the tool initially locates the workpiece edge. While this initial edge-
low-frequency components of the measured force 𝑓 drops below finding is unavoidable, its length is reduced with the SRAM algorithms
50% of the desired force 𝑓𝑑 . compared to ADMIT, as the adaptive admittance tuning automatically
2. Over-force: Over-force error describes when the tool applies ex- sets 𝛾𝑎 ≈ 1 for maximum edge-finding speed. The SRAM-D controller
cessive force, damaging the workpiece and causing increased further reduces the length of this segment due to its feedrate mod-
tool wear. This metric is defined when the low-frequency com- ulation scheme, which lowers the path speed until stable contact is
ponents of the measured force 𝑓 rise above 50% over the desired achieved.
force 𝑓𝑑 . In all registration profiles, the force signal power in the chatter
3. Chatter power: Chatter power evaluates the high-frequency in- region is notably higher in the ADMIT controller compared to the
teraction forces that correlate to degraded surface finish and SRAM and SRAM-D controllers. This difference is due to the adaptive
increased tool wear. It is defined as the band-limited power in admittance tuning in the SRAM and SRAM-D controllers, which allows
the force error signal, within the range of the structural natural them to use high virtual damping for the majority of the test, and
frequencies of the manipulator. only switch to low damping when the state converges at the beginning

13
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Fig. 17. Visualization of feedrate modulation during sample SRAM-D deburring test
from CAM+ registration profile.

of the run. Consequently, the high-frequency components in the force

signal are dampened for most of the test, and are only elevated for
a short period in the beginning. In contrast, ADMIT is forced to use
a constant damping the entire run, tuned with the disparate goals of
being high enough to avoid chatter, but low enough to adapt quickly
to path errors.
Table 4 summarizes the total path distance in contact-loss or over-
force states, and chatter power, averaged over the 60 deburring tests.
The table reiterates the improvements seen in the sample data in
Fig. 16. The ADMIT controller spends significant portions of the de-
burring path in a contact error state—particularly in the less accurate
registration profiles—and exhibited larger chatter power. Contact-loss
and over-force were similar and negligible between all algorithms for
the LASER registration profile, but for CAM, the proposed SRAM-D
controller had 97% lower contact-loss error and 98% lower over-force
compared to the ADMIT controller. Similarly, in the CAM+ profile,
SRAM-D had 99% less contact-loss and 100% less over-force. The
nominal SRAM algorithm had similar, smaller improvements over the
ADMIT controller. Comparisons in the CAM++ profile are not possible
due to ADMIT failing entirely with this registration profile. Chatter
power was mostly independent of registration profile; averaged across
registration profiles, SRAM-D reduced chatter power by 53% compared
to ADMIT and SRAM by 32%. Thus, overall, the SRAM algorithms show
significant improvements over the nominal ADMIT controller in contact
stability and chatter reduction.
Table 4 also reveals that, in general, the proposed improvements
of the SRAM-D controller resulted in better contact stability than the
nominal SRAM algorithm. On average, SRAM-D spent 0.24 mm in
either a contact-loss or overforce state, whereas SRAM spent 0.41 mm—
a 41% improvement. Additionally, it can be qualitatively noted in
Fig. 16 that the force spikes near the large scallops on the left edge of
the part are significantly smaller with the SRAM-D algorithm than with
SRAM. This improved performance can be primarily attributed to the
addition of the novel feedrate modulation scheme from Section 2.3.2.
When the force error increases too high or dips too low, the feedrate
is automatically slowed to allow the force controller to stabilize the
contact.
Fig. 17 visualizes this behavior, plotting the feedrate modulation
in a sample test using the SRAM-D controller and CAM+ registration
profile. For most of the run, the feedrate is kept at the nominal feedrate,
but is reduced at several key points. At the start of the path, the feedrate
is slowed while the tool initially ‘‘finds’’ the workpiece edge, reducing
contact-loss error. Similarly, at the beginning of the first corner, the
feedrate is slowed as the algorithm converges the horizontal placement
Fig. 16. Contact force stability of sample deburring tests for each controller and of the part. Additionally, at many of the sharp corners of the workpiece,
registration profile. The dark blue force 𝑓 is lowpassed at 8 Hz, and the light blue the feedrate is slightly reduced to allow the force controller to stabilize.
force is lowpassed at 30 Hz to show higher frequency chatter power. Dark red dashed These adjustments explain the improvements in contact-loss and over-
lines show the location of the large artificial scallops on the left edge of the part. force error shown in Table 4, as the contact errors are avoided by
giving the controller more time to stabilize when force error increases.

14
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Fig. 18. Comparative images of deburred parts.

Finally, the feedrate is also slowed significantly near the three scallops deburred, whereas large unprocessed edge sections are visible in
on the left edge of the part, allowing the SRAM-D controller to more the ADMIT workpiece from contact-loss error.
effectively remove workpiece imperfections and improve workpiece • Fig. 18d shows the machining quality on straight or continuous
quality. Further discussions on scallop removal are provided in the edges, where the ADMIT algorithm has less difficulty maintaining
following section. stable contact. In these sections, the edge quality is acceptable
The effects of the improved contact stability of the SRAM frame- across all algorithms. However, on the rounded edge, the nibbling
work are also evident in the processed workpieces themselves. Fig. 18 scallops are not entirely removed in the ADMIT algorithm.
presents a selection of annotated images of the processed workpieces, • Fig. 18e displays the same bottom lobe as Fig. 18b from the
comparing the three controllers, a manually deburred part, and a
face view. The profile of the SRAM-D, SRAM, and manual parts
nominal (unprocessed) part as follows:
is accurate; however, the corners of the ADMIT workpiece have
• Fig. 18a shows the edge of the center bottom lobe of sample parts been rounded and modified from the original periphery due to
from the CAM+ registration profile. The SRAM-D, SRAM, and over-force error.
manual parts are deburred properly, while the ADMIT workpiece
has a large unprocessed section from contact-loss error and an The final image, Fig. 18f, compares scallop removal on the left part
improperly removed scallop. edge and is discussed in the following section.
• Similarly, Fig. 18b shows the lower-right lobe of the same parts, We note that, despite the elevated chatter power observed in the
where the SRAM-D and SRAM workpieces are properly deburred, ADMIT controller, the nominal surface finish in the ADMIT work-
but the ADMIT controller has damaged the corner due to over- pieces is relatively smooth, excluding contact-loss and over-force error
force error, compromising the workpiece periphery. The rounded regions. Firstly, this work uses compliant flapwheel tools, which intro-
edge that follows is also improperly deburred on the ADMIT duce additional damping and absorb tool chatter. However, if more
workpiece from contact-loss. rigid tools such as unitized disks or milling bits were used, chatter
• Fig. 18c presents two images from the top-left workpiece corner. would become much more prominent in the surface finish quality. Sec-
In both, the SRAM-D, SRAM, and manual parts are properly ondly, the observed chatter levels and surface finish are a consequence

15
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Table 5 Table 6
Processed scallop size and percent removal. Average tool wear of tested controllers.
Scallop Small Medium Large Ave. ADMIT SRAM SRAM-D
0.32 [0.013’’] 0.64 [0.025’’] 1.27 [0.050’’]
LASER 1.4 mm 1.5 mm (↑ 9%) 1.3 mm (↓ 5%)
SRAM-D <0.1 (↓100%) 0.17 (↓73%) 0.46 (↓64%) 0.23 CAM 2.3 mm 1.7 mm (↓ 25%) 1.1 mm (↓ 50%)
SRAM <0.1 (↓100%) 0.25 (↓61%) 0.71 (↓44%) 0.34 CAM+ 2.0 mm 1.3 mm (↓ 35%) 1.6 mm (↓ 23%)
ADMITa <0.1 (↓100%) 0.29 (↓54%) 0.90 (↓29%) 0.42 CAM++ (FAIL) 1.1 mm (—) 1.3 mm (—)
Manual <0.1 (↓100%) <0.1 (↓100%) <0.1 (↓100%) <0.1 Average 1.9 mm 1.4 mm (↓ 26%) 1.3 mm (↓ 29%)

Data averaged over the 20 tests done with each controller. All measurements in Percentages are given as relative difference from ADMIT. Data averaged over the 5
millimeters. Scallops smaller than 0.1 mm are not possible to measure accurately tests for each controller and registration profile combination.
relative to the workpiece profile.
a
The nominal admittance controller ADMIT failed during the CAM++ registration.
These samples are excluded from the average.
4.5. Tool wear

of the tuning of the ADMIT controller. We selected a nominal damping Tool wear is another key deburring metric, as deburring tooling
of 𝑏𝑎 = 0.75 Ns/mm as an optimal trade-off between edge-following represents a considerable ongoing cost for any machining workcell.
and contact stability. Reducing this damping further improves edge- Therefore, it is essential in a robotic deburring setup to take steps to
tracking and reduces over-force and contact-loss error, but leads to an minimize tool wear. As discussed in Section 2.2.2, tool wear primarily
unstable contact in which chatter marks did become noticeable, and correlates to the applied force, with larger tool forces causing exponen-
tool wear significantly increased. Conversely, increasing the nominal tially greater wear. Consequently, the improved contact stability and
damping reduced chatter power but increased contact errors. Thus,
reduced over-force error observed in the SRAM methods are expected
without the adaptive tuning of the SRAM framework, achieving perfect
to decrease wear compared to a nominal admittance controller.
tuning with a nominal admittance controller is challenging.
Tool wear is challenging to measure after testing with calipers, since
4.4. Scallop removal the tool changes shape while being rotated by the spindle, and the flaps
deflect under light loads when using calipers. Instead, tool size can
In addition to contouring the part with constant force, a deburring be more accurately measured from the deburring test data itself. By
process should also identify and remove large workpiece imperfections observing the tool position throughout the run, and the corresponding
such as burrs, scallops, or index tabs. In the current tests, workpieces contact forces, the contact model described in Section 2.2.3 can be used
have three artificial scallops added to the left edge of the part, as
to accurately estimate the tool radius throughout the tests. These tool
shown in Fig. 7, with small (0.32 mm), medium (0.64 mm), and
wear estimates are summarized in Table 6, averaging results across
large (1.27 mm) sizes, respectively. These repeatable scallops allow
controller performance to be quantified in removing large workpiece the 60 deburring tests. The results generally show reduced wear by
imperfections. The large scallop is bigger than what would be encoun- the SRAM and SRAM-D algorithms, particularly in the less accurate
tered in typical deburring applications, but is included to test the limits registration profiles. For LASER, the tool wear was similar across all
of the controllers. controllers; however, for CAM, SRAM reduced tool wear by 25% and
The height of the artificial scallops on the processed workpieces SRAM-D by 50%. In CAM+, SRAM reduced tool wear by 35%, while
is measured using high-precision calipers and summarized in Table 5, SRAM-D by 23%. We note that flapwheel tools are not manufactured
averaging scallop sizes over the 20 tests done with each controller. somewhat inconsistently, resulting in variations in wear run-to-run.
The results show that the SRAM-D algorithm removes significantly Therefore, the results in Table 6 are somewhat noisy. However, the
more of the artificial scallops than both SRAM or ADMIT, thanks to
overall trend is clear: the improved contact stability of the SRAM frame-
the novel spring constant and feedrate modulation schemes introduced
work allows for a reduction in tool wear. Averaged between registration
in Section 2.3. All controllers completely remove the small scallop.
However, with the medium scallop, the SRAM-D controller removed profiles, SRAM-D offered a 4% reduction in tool wear compared to
73%, whereas SRAM removed only 61% and ADMIT 54%. Similarly, SRAM, and a significant 29% reduction compared to ADMIT.
with the large scallop, SRAM-D removed 64%, whereas SRAM removed It should also be noted that while SRAM-D, and to a lesser ex-
only 44%, and ADMIT 29%. Thus, on average, SRAM-D removed the tent SRAM, have lower tool wear, they also perform more material
medium and large imperfections 45% more effectively than ADMIT and removal. The scallops on the left side of the part were significantly more
34% more effectively than SRAM. In all cases, the manually deburred smoothed with SRAM-D and SRAM, whereas ADMIT merely lightly
part was perfect, as the human operator continued processing the large passed over them. Similarly, other scallops around the part were more
scallops until they were fully removed. effectively removed in the SRAM methods. Thus, with both SRAM
Fig. 18f shows sample processed workpieces from each of the tested
controllers, tool wear is lowered, reducing costs—while also performing
controllers. All methods removed the bulk of the imperfection on the
more thorough deburring and processing of the part.
small scallop, but the ADMIT controller left behind a rough finish. For
the medium and large scallops, none of the robotic controllers removed In these tests, our objective is to quantify the relative improvements
the imperfections entirely. However, SRAM-D visibly removes much offered by the SRAM frameworks in deburring. We do not claim that the
more than the other controllers. Incomplete scallop removal by the numbers in Table 6 represent any absolute reduction in wear compared
ADMIT algorithm can also be seen in Fig. 18a and d, as described to other works. Several optimizations could further reduce tool wear in
earlier. this application. For example, dithering the tool along its axis during
Thus, the performance of the SRAM-D controller is significantly the deburring would wear the tool more gradually—particularly given
improved over the nominal controllers as a result of the proposed the thin thickness of the part. However, such adjustments would also
spring constant and feedrate modulation schemes introduced in Sec-
make it challenging to compare wear run-to-run and are not performed
tion 2.3. The spring constant modulation penalizes deflections of the
in the current tests.
force controller away from the part, forcing the tool to pass through
the imperfection. Simultaneously, the feedrate modulation slows down Overall, the tool wear is significantly reduced with the SRAM meth-
the toolpath in response to the increased interaction forces, spending ods compared to the nominal ADMIT controller, representing a signifi-
more time on the imperfection and preventing excessive workpiece-tool cant cost saving in industrial implementations through reduced ongoing
contact forces. tooling costs.

16
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Table 7 deburring can typically only be performed during work hours, and
Average cycle times of tested registration methods.
workers require rest and breaks throughout the day, while a robotic
Scanning Planning Deburr Totala manipulator does not. As such, direct comparisons of worker and robot
LASER 47 s 30 s 140 s 217 s efficiency are difficult. In situations where the high quality of manual
CAM – 33 s 140 s 173 s
deburring is necessary, it is also possible for manufacturers to perform
Manual – – 209 s 209 s
the bulk of deburring using an automated workcell and have a human
Data averaged over 12 tests done with LASER registration profile, 48 tests done with worker inspect and touch up the processed parts more quickly than
CAM, CAM+, or CAM++ profile, and 5 manual tests.
a
fully deburring the part themselves.
Assumes perfect sequencing of operations—in reality, laser localization required
additional time due to imperfect sequencing of operations.
5. Conclusion

This paper demonstrated the adaptation and application of the

4.6. Cycle times
Simultaneous Registration and Machining framework for robotic de-
burring, resulting in improved machining quality, reduced tool wear,
Reducing overall workcell cycle time is crucial to efficiency. The
and the ability to use faster, less costly, yet less precise registration
cycle times of the registration approaches investigated in this paper
methods. We propose the SRAM-D controller, which incorporates two
are summarized in Table 7. The CAM registration is the fastest method,
improvements over the nominal SRAM method: a spring constant mod-
with LASER requiring an additional 47 s to scan the part. Both require
ulation scheme and a feedrate modulation scheme. Together, these
about 30 s for planning and 140 s for deburring, resulting in a total
improvements allow for more effective removal of scallops and burrs,
cycle time of 217 s for LASER and 173 s for CAM. Thus, camera
improve contact stability, and optimize cycle times. The application
localization improves workcell efficiency by approximately 27%.
of the method is described, including a calibration procedure for the
However, there are some additional points to consider. First, the
deburring tool model and guidelines for tuning the SRAM-D controller.
numbers in Table 7 assume perfect sequencing of operations. In re- SRAM-D is implemented in a custom robotic deburring workcell and
ality, laser localization took significantly longer than reported with rigorously validated in extensive experimental testing, against nominal
our system, due to additional time spent on startup procedures, data SRAM and nominal admittance controllers under varying levels of
processing, path verification, and more. In an industrial setup, this registration uncertainty.
sequencing could be done more efficiently, but may not be reduced Test results show that the SRAM-D controller significantly outper-
fully to the levels reported in the table. Second, in the current testing, forms the nominal SRAM and ADMIT controllers in deburring. Path
workpieces were relatively small. As workpiece size increases, scanning error was minimized and contact stability enhanced, improving work-
area and scanning time increase quadratically. Point cloud sizes also piece quality. Contact-loss and over-force error was reduced by 99%
increase with scanning area, slowing the localization and planning compared to ADMIT, and 41% compared to the nominal SRAM con-
steps. Conversely, planning time for the camera system remains con- troller. High-frequency chatter power was reduced by 53% compared
stant, regardless of workpiece size. Lastly, laser scanner systems are to the nominal ADMIT controller and 32% compared to the nominal
considerably more expensive than camera systems—in our setup, the SRAM. The proposed additions increased burr removal by 45% com-
laser scanner system cost over an order of magnitude more than the pared to ADMIT and 34% compared to the nominal SRAM. Improved
camera system. contact stability in SRAM-D also led to a 29% reduction in tool wear
Therefore, camera localization represents the ideal choice for work- compared to ADMIT and a 4% reduction compared to SRAM—despite
piece registration, reducing costs and cycle times—as long as its re- SRAM-D also machining parts and removing burrs more effectively.
duced accuracy can be tolerated. The proposed SRAM-D controller The enhanced SRAM-D controller is proposed for superior contact
allows such errors to be handled without adversely affecting cycle stability and workpiece quality in robotic deburring, with more ef-
times or workpiece quality. Rather, contact stability, workpiece quality, ficient removal of burrs and decreased tool wear. Lower tool wear
and tool wear are improved, as discussed in the previous sections. reduces ongoing tooling costs for workcells, while the ability to handle
Moreover, the proposed SRAM-D method allows registration errors lower-quality registration enables the use of faster and more cost-
from other sources, such as a manipulator without kinematic cali- effective registration methods, such as vision systems, ultimately re-
bration (typical in industrial settings due to cost and complexity), a ducing workcell expenses and increasing workcell efficiency.
lower repeatability robotic manipulator, a lower tolerance end-effector
design, and more. This allowance enables simpler design of workcells, CRediT authorship contribution statement
with less costly components, and allows complex and costly kinematic
calibration procedures to be avoided. Steffan Lloyd: Conceptualization, Data curation, Formal analysis,
The SRAM-D controller also implements the novel real-time feedrate Investigation, Methodology, Software, Validation, Visualization, Writ-
modulation scheme proposed in Section 2.3.2, allowing for further ing – original draft. Rishad A. Irani: Funding acquisition, Project
reductions in cycle times. For a given application, the maximum fee- administration, Supervision, Writing – review & editing. Mojtaba Ah-
drate of the nominal SRAM and ADMIT controllers is determined madi: Funding acquisition, Project administration, Supervision, Writ-
by a speed that can remove the largest burrs or scallops. However, ing – review & editing.
SRAM-D automatically slows near large workpiece imperfections to
remove them more effectively. As a result, the nominal feedrate can be Declaration of competing interest
increased accordingly. Optimizing feedrates for the complete removal
of the largest imperfections depends on the application and was not The authors declare the following financial interests/personal rela-
feasible to investigate in the current study. Nonetheless, it offers a tionships which may be considered as potential competing interests:
promising avenue to further reduce cycle times and enhance efficiency. Steffan Lloyd reports financial support was provided by Natural Sci-
Manual deburring time is also quantified in Table 7 as 209 s, ences and Engineering Research Council of Canada. Rishad Irani reports
roughly the same as the LASER localization. The quality of the manual financial support was provided by Natural Sciences and Engineering
deburring was also high, since workers could inspect and correct the Research Council of Canada. Mojtaba Ahmadi reports financial support
part during processing. However, manual deburring has numerous was provided by Natural Sciences and Engineering Research Council
other issues, as mentioned in the introduction, such as health risks, of Canada. Robotic deburring research project (PI Mojtaba Ahmadi)
high turnover, and low rates of skill transfer. Furthermore, manual reports financial support was provided by Arnprior Aerospace Inc.

17
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

Data availability [21] E. Villagrossi, C. Cenati, N. Pedrocchi, M. Beschi, L. Molinari Tosatti, Flexible
robot-based cast iron deburring cell for small batch production using single-
point laser sensor, Int. J. Adv. Manuf. Technol. 92 (2017) 1425–1438, http:
The authors do not have permission to share data.
//dx.doi.org/10.1007/s00170-017-0232-2.
[22] J.R. Diaz Posada, U. Schneider, S. Pidan, M. Geravand, P. Stelzer, A. Verl,
Acknowledgments High accurate robotic drilling with external sensor and compliance model-
based compensation, in: Proc. - IEEE Int. Conf. Robot. Autom., IEEE, 2016, pp.
3901–3907, http://dx.doi.org/10.1109/ICRA.2016.7487579.
This work was supported in part by the Natural Sciences and Engi- [23] H. Kosler, U. Pavlovčič, M. Jezeršek, J. Možina, Adaptive robotic deburring of
neering Research Council of Canada (NSERC) [grant numbers RGPIN- die-cast parts with position and orientation measurements using a 3D laser-
2022-04934, RGPIN-2021-04207, and CRDPJ 514258-17], and indus- triangulation sensor, Stroj. Vestnik/J. Mech. Eng. 62 (2016) 207–212, http:
//dx.doi.org/10.5545/sv-jme.2015.3227.
trial partner Arnprior Aerospace Inc. Support for Steffan Lloyd was
[24] G. Xiong, Z.L. Li, Y. Ding, L.M. Zhu, A closed-loop error compensation method
provided by NSERC CGS-D. for robotic flank milling, Robot. Comput.-Integr. Manuf. 63 (2020) 101928,
http://dx.doi.org/10.1016/j.rcim.2019.101928.
References [25] M.-H. Liu, H. Tolle, Force-controlled robotic deburring, IFAC Proc. Vol. 26 (1993)
855–858, http://dx.doi.org/10.1016/s1474-6670(17)48592-2.
[26] K.M. Murphy, R.J. Norcross, F.M. Proctor, CAD directed robotic deburring, in:
[1] L.K. Gillespie, Deburring and Edge Finishing Handbook, Society of Manufacturing
Proc. Second Int. Symp. Robot. Manuf. Res. Educ. Appl. Albuquerque, NM, 1988.
Engineers, 1999. [27] N. Asakawa, K. Toda, Y. Takeuchi, Automation of chamfering by an industrial
[2] F. Leali, M. Pellicciari, F. Pini, G. Berselli, A. Vergnano, An offline programming robot; for the case of hole on free-curved surface, Robot. Comput.-Integr. Manuf.
method for the robotic deburring of aerospace components, in: Commun. 18 (2002) 379–385.
Comput. Inf. Sci., vol. 371, 2013, pp. 1–13, http://dx.doi.org/10.1007/978-3- [28] H.C. Song, J.B. Song, Precision robotic deburring based on force control for
642-39223-8_1. arbitrarily shaped workpiece using CAD model matching, Int. J. Precis. Eng.
[3] K. Stouffer, R. Russell, R. Archacki, T. Engel, R. Dansereau, A. Grot, Advanced Manuf. 14 (2013) 85–91, http://dx.doi.org/10.1007/s12541-013-0013-2.
Deburring and Chamfering System (ADACS): Final Report, Tech. Rep. September, [29] M. Jinno, M. Uenohara, J. Oaki, K. Tatsuno, Teaching-less robot system for
National Institute of Standards and Technology, 1996. finishing workpieces of various shapes using force control and computer vision,
[4] H. Zhang, J. Wang, G. Zhang, Z. Gan, Z. Pan, H. Cui, Z. Zhu, Machining IEEE Int. Conf. Intell. Robot. Syst. 1 (1999) 573–578, http://dx.doi.org/10.1109/
with flexible manipulator: Toward improving robotic machining performance, iros.1999.813065.
in: Proceedings, 2005 IEEE/ASME Int. Conf. Adv. Intell. Mechatronics, 2, IEEE, [30] A. Kuss, M. Drust, A. Verl, Detection of workpiece shape deviations for tool
2005, pp. 1127–1132, http://dx.doi.org/10.1109/AIM.2005.1511161. path adaptation in robotic deburring systems, Procedia CIRP 57 (2016) 545–550,
[5] L.K. Gillespie, Hand Deburring: Increasing Shop Productivity, Society of http://dx.doi.org/10.1016/j.procir.2016.11.094.
Manufacturing Engineers, Dearborn, Michigan, 2003, pp. 2–4. [31] B. Mei, Z. Liang, W. Zhu, Y. Ke, Positioning variation synthesis for an automated
[6] Deburring airframe components, Aircr. Eng. Aerosp. Technol. 60 (3) (1988) drilling system in wing assembly, Robot. Comput.-Integr. Manuf. 67 (2021)
29–30, http://dx.doi.org/10.1108/eb036591. 102044, http://dx.doi.org/10.1016/j.rcim.2020.102044.
[7] J. Aurich, Ergebnisbericht der Untersuchung zur Beherrschung der Sauberkeit [32] Z. Lai, R. Xiong, H. Wu, Y. Guan, Integration of visual information and robot
von zerspanend hergestellten Bauteilen, Tech. rep., Ergebnisworkshop; Lehrstuhl offline programming system for improving automatic deburring process, in: 2018
für Fertigungstechnik und Betriebsorganisation, Technische Universität Kaiser- IEEE Int. Conf. Robot. Biomimetics, ROBIO 2018, IEEE, 2018, pp. 1132–1137,
slautern, 2006. http://dx.doi.org/10.1109/ROBIO.2018.8665148.
[8] B.M. Kramer, J.J. Bausch, R.L. Gott, D.M. Dombrowski, Robotic deburring, [33] F. Leo Princely, T. Selvaraj, Vision assisted robotic deburring of edge burrs in cast
Robot. Comput.-Integr. Manuf. 1 (3–4) (1984) 365–374, http://dx.doi.org/10. parts, Procedia Eng. 97 (2014) 1906–1914, http://dx.doi.org/10.1016/j.proeng.
1016/0736-5845(84)90026-7. 2014.12.344.
[9] International Federation of Robotics, World Robotics 2019, Statistical Yearbook, [34] T.F. Lu, G.C. Lin, An on-line relative position and orientation error calibration
Tech. rep., 2019. methodology for workcell robot operations, Robot. Comput.-Integr. Manuf. 13
[10] A. Karim, A. Verl, Challenges and obstacles in robot-machining, in: 2013 44th (1997) 89–99, http://dx.doi.org/10.1016/S0736-5845(96)00029-4.
Int. Symp. Robot. ISR 2013, IEEE, 2013, pp. 16–19, http://dx.doi.org/10.1109/ [35] M. Dinham, G. Fang, Autonomous weld seam identification and localisation using
ISR.2013.6695731. eye-in-hand stereo vision for robotic arc welding, Robot. Comput.-Integr. Manuf.
[11] I.F. Onstein, O. Semeniuta, M. Bjerkeng, Deburring using robot manipulators: 29 (2013) 288–301, http://dx.doi.org/10.1016/j.rcim.2013.01.004.
A review, in: 2020 3rd Int. Symp. Small-Scale Intell. Manuf. Syst. SIMS 2020, [36] B.-S. Ryuh, G.R. Pennock, Robot automation systems for deburring, in: L.K. Huat
2020, http://dx.doi.org/10.1109/SIMS49386.2020.9121490. (Ed.), Ind. Robot. Program. Simul. Appl., Pro Literatur Verlag, Germany / ARS,
[12] A. Verl, A. Valente, S. Melkote, C. Brecher, E. Ozturk, L.T. Tunc, Robots Austria, 2006, http://dx.doi.org/10.5772/4919.
[37] M.H. Liu, Force-controlled fuzzy-logic-based robotic deburring, Control Eng.
in machining, CIRP Ann. 68 (2019) 799–822, http://dx.doi.org/10.1016/j.cirp.
Pract. 3 (1995) 189–201, http://dx.doi.org/10.1016/0967-0661(94)00076-S.
2019.05.009.
[38] N. Hogan, Impedance control: An approach to manipulation, J. Dyn. Syst. Meas.
[13] M. Jonsson, A. Stolt, A. Robertsson, S. von Gegerfelt, K. Nilsson, On force
Control 107 (1985) 1–24, http://dx.doi.org/10.1115/1.3140702.
control for assembly and deburring of castings, Prod. Eng. 7 (2013) 351–360,
[39] N. Hogan, Impedance control applied to automated deburring, in: Adapt. Learn.
http://dx.doi.org/10.1007/s11740-013-0459-1.
Syst., Springer US, Boston, MA, 1986, pp. 359–366, http://dx.doi.org/10.1007/
[14] S. Lloyd, R. Irani, M. Ahmadi, A framework for simultaneous workpiece reg-
978-1-4757-1895-9_25.
istration in robotic machining applications, in: 2023 IEEE Int. Conf. Robot.
[40] H. Kazerooni, Robotic deburring of two-dimensional parts with unknown ge-
Autom., IEEE, London, UK., 2023, pp. 5249–5255, http://dx.doi.org/10.1109/
ometry, J. Manuf. Syst. 7 (1988) 329–338, http://dx.doi.org/10.1016/0278-
ICRA48891.2023.10160445.
6125(88)90043-X.
[15] S. Lloyd, R.A. Irani, M. Ahmadi, Improved accuracy and contact stability in [41] M.S. Ali, M.N. Noori, J. Turi, Automatic deburring utilizing a real-time
robotic contouring with simultaneous registration and machining, IEEE Trans. impedance control strategy, Comput. Struct. 46 (3) (1993) 561–571, http://dx.
Control Syst. Technol. (2024) http://dx.doi.org/10.1109/TCST.2023.3348749, In doi.org/10.1016/0045-7949(93)90225-3.
press. [42] S. Lloyd, R. Irani, M. Ahmadi, Application of pseudo-symbolic dynamic modeling
[16] J. Posada, R. Diaz, S. Kumar, A. Kuss, U. Schneider, M. Drust, T. Dietz, A. Verl, (PSDM) in the modeling and calibration of a 6-DOF articulated robot, Model.
Automatic programming and control for robotic deburring, in: Proc. ISR 2016 Estim. Control Conf. 55 (37) (2022) 74–80, http://dx.doi.org/10.1016/j.ifacol.
47st Int. Symp. Robot., VDE, 2016. 2022.11.164.
[17] A. Joubair, I.A. Bonev, Kinematic calibration of a six-axis serial robot using [43] D.E. Whitney, Force feedback control of manipulator fine motions, J. Dyn. Syst.
distance and sphere constraints, Int. J. Adv. Manuf. Technol. 77 (2015) 515–523, Meas. Control. Trans. ASME 99 (2) (1977) 91–97, http://dx.doi.org/10.1115/1.
http://dx.doi.org/10.1007/s00170-014-6448-5. 3427095.
[18] T.S. Lembono, F. Suarez-Ruiz, Q.-C. Pham, SCALAR: Simultaneous calibration [44] J.E. Solanes, L. Gracia, P. Muñoz-Benavent, A. Esparza, J. Valls Miro, J. Tornero,
of 2-D laser and robot kinematic parameters using planarity and distance Adaptive robust control and admittance control for contact-driven robotic surface
constraints, IEEE Trans. Autom. Sci. Eng. 16 (2019) 1971–1979, http://dx.doi. conditioning, Robot. Comput.-Integr. Manuf. 54 (2018) 115–132, http://dx.doi.
org/10.1109/TASE.2019.2918141. org/10.1016/j.rcim.2018.05.003.
[19] M. Dehghani, R.A. McKenzie, R.A. Irani, M. Ahmadi, Robot-mounted sensing and [45] F.Y. Hsu, L.C. Fu, Intelligent robot deburring using adaptive fuzzy hybrid
local calibration for high-accuracy manufacturing, Robot. Comput.-Integr. Manuf. position/force control, IEEE Trans. Robot. Autom. 16 (4) (2000) 325–335, http:
79 (2022) 102429, http://dx.doi.org/10.1016/j.rcim.2022.102429. //dx.doi.org/10.1109/70.864223.
[20] J. Hu, P.R. Pagilla, Dual-edge robotic gear chamfering with registration error [46] X.L. Wang, Y. Wang, Y.N. Xue, An adaptive algorithm for robotic deburring
compensation, Robot. Comput.-Integr. Manuf. 69 (2021) 102082, http://dx.doi. based on impedance control, IEEE Int. Symp. Ind. Electron. 1 (2006) 262–266,
org/10.1016/j.rcim.2020.102082. http://dx.doi.org/10.1109/ISIE.2006.295603.

18
S. Lloyd et al. Robotics and Computer-Integrated Manufacturing 88 (2024) 102733

[47] C. Ott, R. Mukherjee, Y. Nakamura, Unified impedance and admittance control, [50] W. Rudin, Principles of Mathematical Analysis, 3rd editio, McGraw Hill, 1976,
Proc. - IEEE Int. Conf. Robot. Autom. (2010) 554–561, http://dx.doi.org/10. pp. 223–228.
1109/ROBOT.2010.5509861. [51] S. Lloyd, R.A. Irani, M. Ahmadi, Fast and robust inverse kinematics of serial
[48] K. Schützer, E. Abele, S. Güth, Simulation-based deburring tool and process robots using Halley’s method, IEEE Trans. Robot. 38 (5) (2022) 2768–2780,
development, CIRP Ann. - Manuf. Technol. 64 (1) (2015) 357–360, http://dx. http://dx.doi.org/10.1109/TRO.2022.3162954.
doi.org/10.1016/j.cirp.2015.04.099. [52] S. Lloyd, R. Irani, M. Ahmadi, A numeric derivation for fast regressive modeling
[49] W.R. MacMillan, R.A. Irani, M. Ahmadi, Planar image-space trajectory planning of manipulator dynamics, Mech. Mach. Theory 156 (2021) 104149, http://dx.
algorithm for contour following in robotic machining, CIRP J. Manuf. Sci. doi.org/10.1016/j.mechmachtheory.2020.104149.
Technol. 42 (2023) 1–11, http://dx.doi.org/10.1016/j.cirpj.2023.01.005. [53] W. Macmillan, Contour Following and Trajectory Planning for Robotic Deburring
(Ph.D. thesis), Carleton University, 2022.

19