42 IEEE TRANSACTIONS ON ROBOTICS, VOL.

41, 2025

Stability Criterion and Stability Enhancement for a

Thruster-Assisted Underwater Hexapod Robot
Lepeng Chen , Rongxin Cui , Member, IEEE, Weisheng Yan , Chenguang Yang , Fellow, IEEE,
Zhijun Li , Fellow, IEEE, Hui Xu , and Haitao Yu

Abstract—The stability criterion is critical for the design of

legged robots’ motion planning and control algorithms. If these
algorithms cannot theoretically ensure legged robots’ stability,
we need many trials to identify suitable parameters for stable
locomotion. However, most existing stability criteria are tailored
to robots driven solely by legs and cannot be applied to thruster-
assisted legged robots. Here, we propose a stability criterion for a
thruster-assisted underwater hexapod robot by finding maximum
and minimum allowable thruster forces and comparing them with
the current thrusts to check its stability. On this basis, we propose
a method to increase the robot’s stability margin by adjusting
the value of thrusts. This process is called stability enhancement.
Fig. 1. Underwater hexapod robot has six C-shaped legs and eight thrusters
The criterion uses the optimization method to transform multiple (four lateral and four vertical thrusters). The robot body and legs are connected
variables such as attitude, velocity, acceleration of the robot body, by hip joints [4], [5]. Legs and thrusters simultaneously drive the robot, which
and the angle and angular velocity of leg joints into one kind of can walk on the inclined and rough underwater terrains.
variable (thrust) to judge the stability directly. In addition, the
stability enhancement method is straightforward to implement
because it only needs to adjust the thrusts. These provide insights
into how multiclass forces such as inertia force, fluid force, thrust, many trials are needed to find the feasible parameters for stable
gravity, and buoyancy affect the robot’s stability. locomotion [3]. Thus, a clear stability criterion is essential, one
Index Terms—Stability criterion, stability enhancement, thru- that not only evaluates the robot’s stability but also supports
ster-assisted underwater hexapod robot. the development of stability constraints for designing robust
planning and control algorithms.
To perform some meticulous small-range operations on the
I. INTRODUCTION underwater structure’s surface, the robot should be able to walk
on it. Then, we develop a thruster-assisted underwater hexapod
TABILITY is a critical issue for underwater legged robots
S because they may suffer from multiclass disturbances while
walking, such as waves, currents, towing force of cable, and
robot (see Fig. 1) [4], [5]. Unlike underwater robots driven
by legs that only walk on the seabed, our robot can walk on
structures with various inclinations and surface roughness.
uncertain leg-terrain interaction force [1], [2]. The robot is stable Most existing stability criteria are with robots only driven by
if it can recover to the preset motion state by itself under the legs, and they cannot be applied to our robot because it is driven
disturbances. Otherwise, it is unstable. In addition, if motion by legs and thrusters simultaneously. Moreover, it is challenging
planning and control algorithms cannot theoretically ensure to identify how the other multiclass forces, such as inertia force,
stability, they cannot be applied to other robots directly, and fluid force, gravity, and buoyancy, affect the robot’s stability and
enhance it by regulating thrusts.
Received 20 September 2024; accepted 12 October 2024. Date of publi-
cation 6 November 2024; date of current version 6 December 2024. This
work was supported in part by the National Natural Science Foundation of
A. Related Works
China (NSFC) under Grant U22A2066, Grant U21B2047, Grant 61733014, The dynamic model can be used to predict the legged robot’s
and Grant 62103182, and in part by the Doctorate Foundation of Northwestern
Polytechnical University under Grant CX201904. This article was recommended
motion by establishing the relationship between multiclass
for publication by Associate Editor Y. Liu and Editor S. Behnke upon evaluation forces and the motion states. Knowing the future motions con-
of the reviewers’ comments. (Corresponding author: Rongxin Cui.) tributes to evaluating the robot’s stability. The model mainly
Lepeng Chen, Rongxin Cui, Weisheng Yan, Hui Xu, and Haitao Yu are with consists of two types: inverted pendulum model [6], [7], [8], [9]
the School of Marine Science and Technology, Northwestern Polytechnical and force balancing model [10], [11], [12], [13]. The former is
University, Xi’an 710072, China (e-mail: r.cui@nwpu.edu.cn).
Chenguang Yang is with the Department of Computer Science, University of established by simplifying the robot as an inverted pendulum,
Liverpool, L69 3BX Liverpool, U.K.. and the latter is an equilibrium equation to balance multiclass
Zhijun Li is with the School of Mechanical Engineering, and Translational Re- forces, e.g., leg-terrain interaction forces, inertia forces, and
search Center, Shanghai Yang Zhi Rehabilitation Hospital (Shanghai Sunshine other forces.
Rehabilitation Center), Tongji University, Shanghai 200070, China.
This article has supplementary downloadable material available at
Inspired by the spring-loaded inverted pendulum (SLIP)
https://doi.org/10.1109/TRO.2024.3492374, provided by the authors. model for ground legged robots, Calisti et al. [6] proposed a
Digital Object Identifier 10.1109/TRO.2024.3492374 new dynamic model, called underwater spring-loaded inverted

1941-0468 © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 43

pendulum (U-SLIP), for underwater legged robots. On this basis, Previously, considering the alternating support character, we
Picardi et al. [7] presented a three-dimensional (3-D) U-SLIP proposed a stability criterion for our robot in [23], extending the
model that includes an underwater legged robot’s lateral dy- concept of the capture point to a capture domain and determining
namic motion. Lin et al. [8], [9] presented a rolling inverted stability based on the relationship between the capture and
pendulum model for a ground hexapod robot by using the support domains. However, this criterion is only applicable to
Lagrange method and incorporating the rolling character caused flat terrain walking. Besides, it requires both motion states and
by C-shaped legs. These elegant models can only describe the thrust values for stability evaluation, complicating the design of
motion of walking on flat terrain and need an assumption that stability enhancement methods.
the leg-terrain friction should be large enough. According to the difference in principle, the stability en-
Several attractive force balancing models are developed hancement methods can be divided into three categories: robotic
in [10], [11], and [12] for ground quadruped robots, that relax functional mechanism designing [24], [25], [26], [27], supernu-
the above flat terrain and sufficient friction assumptions. Calisti merary tail adding [28], [29], [30], [31], [32], and inner motion
et al. [13] presented a force balancing model for a novel octopus- planning and controlling [33], [34], [35], [36], [37], [38].
inspired underwater robot by considering leg-terrain interaction The first category focuses on modifying robots’ geometri-
forces and drags arising from the sculling motion of the legs. cal and physical properties to buffer the leg-terrain impact or
These models apply only to ground robots with point-contact improve the self-stabilizing capability. Chellapurath et al. [24]
legs. Thus, we need to establish a force balancing model for our disclosed how leg configuration, net weight, and the nature of
robot by considering its unique driving manner and fluid forces. the substrate affect the passive station-keeping performance of
Several insightful stability criteria are presented for legged an underwater legged robot and found that the performance
robots and can be divided into three categories: zero-tilting improves with an increase in net weight, a decrease in height, and
moment point (ZMP) criterion [14], [15], Poincare return map an increase in width. In [25] and [26], two novel self-adaptive
criterion [3], [16], [17], and force balancing criterion [18], [19], toe mechanisms are presented to possess cat claws’ buffer and
[20], [21]. hold performance.
Vukobratovi et al. [14], [15] initially presented the ZMP crite- The second category involves using the drag or inertia force
rion, which checks the stability by judging whether the ZMP lies generated by the added tail to enhance stability. Zhang et al. [28]
within the support polygon. It has been widely used to evaluate presented a new algorithm to alter the system center of gravity by
legged robots’ stability because of its clear physical meaning and adjusting the joint angles of a three DoF manipulator, enhancing
easy-use advantage. Note that this criterion assumes that contact the ability to prevent tip over. Inspired by basilisk lizards, Kim
terrain is flat and leg-terrain friction is adequate. In addition, the et al. [29] developed an interesting water-running robot that uses
ZMP criterion is a constraint to prevent the robot from tipping a two DoF tail to provide drag for enhancing the stability of
over. It only uses the support legs’ rotating information, ignoring yaw and pitch motion. In [30], an elegant optimal controller is
the swing legs’ information. That said, it overlooks the robots’ proposed to steer the joints of the added tail to compensate for
alternating support character. the neural delay dynamics for enhancing the human stance’s
The Poincare return map is a tool to transform the stability stability. Using a deep reinforcement learning method, Huang
of a periodic orbit to a fixed point’s stability [22], which can et al. [31] created a policy to automatically adjust the behaviors
be used to judge periodically walking robots’ stability. Grizzle of the tail to maintain a quadruped robot’s stability.
et al. [16], [17] extended this criterion to a hybrid dynamics The third category relies on the stability criterion or bioinspi-
with impulse effects and applied it to judge the stability of a ration to determine control inputs and plan the robot’s desired
three degree-of-freedom (DoF) humanoid robot with leg-terrain motion. In [33] and [34], the ZMP criterion is used to construct
impacts. Fu et al. [3] presented a section-map stability criterion stability constraints integrated into the trajectory optimization
to check the stability of nonperiodic walking robots. framework to obtain optimal and stable motions for legged
The force balancing criterion relies on the force-balance prin- robots. Choi et al. [35] presented a new method to form the
ciple to check whether the robot can tip over or slip. Hirukawa regions of attraction for underactuated legged robots, and it
et al. [18] presented a novel stability criterion for legged robots, facilitates the design of the controller for stability enhancement.
which checks if the sum of the gravity and inertia wrench lies Using adaptive central pattern generators, Mehr et al. [36]
within the region constructed by noncoplanar contact points. presented a novel motion planner to adjust the upper body’s
Although more universal than the ZMP criterion, it is still position for ensuring a lower limb exoskeleton’s postural sta-
based on the sufficient friction assumption. Caron et al. [19] bility. Imitating animals’ natural gaits, Humphreys et al. [38]
considered friction and linearized pendulum constraints and presented a gait transition method for locomotion proficiency
presented a more necessary and sufficient criterion to check the enhancing.
multicontact stability. Using the static force balancing model Despite the abovementioned methods achieving impressive
and optimization algorithm, Audren et al. [21] constructed a results in stability enhancement, they are mainly applicable to
robust 3-D region that consists of all feasible center of mass robots that allow the addition of tails or robots only driven by
(CoM) positions by suppressing bounded CoM accelerations. legs. They are challenging to apply to legged robots simultane-
If the CoM position remains within this region, the robot is ously driven by legs and thrusters.
stable. These interesting stability criteria are the only necessary
conditions for stable locomotion. The robot may still be stable
due to its alternating support character when these conditions are B. Proposed Approach and Contribution
violated. In this article, we establish a force balancing model and use
In addition, the abovementioned criteria are limited to robots the optimization method to find a feasible set of thrusts that
that are only driven by legs and cannot be applied to our robot. allows for a certain degree of tip over and can also prevent our

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 44 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

Fig. 2. Framework of the stability criterion and stability enhancement method. (a) We establish a force balancing model for our robot. (b) We present a new
method to define three types of walking motions by relying on the static and nonstatic force balancing model and fuse the alternating support character by relaxing
the model. (c) We use the optimization method to find the allowable maximum and minimum vertical thrusts and propose a stability criterion by checking whether
the robot’s vertical thrust belongs to the optimized stable region. Then, we present a stability enhancement method that changes the robot stability margin by
adjusting the value of the vertical thrust. (d) We provide the expression of model relaxations using finite-time control theory to derive the maximum and minimum
forces/moments. Then, we use stricter friction constraints to suppress the effects caused by the model uncertainties. Finally, we present a 6-D stability criterion
and stability enhancement method. (a) Force balancing model. (b) Fuse alternating support character. (c) Vertical thrust based stability criterion and enhancement.
(d) 6-D thrust stability criterion and enhancement.

robot from slipping. Using the allowable vertical thrust and 6-D leg-terrain frictional constraint. Finally, all related criteria do not
thrust vector, we present two solutions to check and enhance consider the joint’s torque limit. In summary, our criterion can
the robot’s stability. Fig. 2 shows the framework of the stability compensate for the deficiencies of the related criteria above. In
criterion and enhancement method. Table I shows the summary addition, once the ZMP position is not in the support region of
of related criteria and our criterion from the aspects of the used a legged robot, the ZMP criterion indicates that the robot will
dynamic model, alternating support character, applied range tip over, part of the legs will lift off the ground and then their
for rough terrain, and whether to consider leg-terrain frictional contact forces will become zero. The force balancing criterion
constraint and joint torque limit. Apart from the Poincare return judges the robot’s stability by checking feasible contact forces’
map criterion, the other related criteria do not fuse the alternating existence. Both criteria use the force-balance principle to check
support character. In addition, except for the force balancing whether existing the feasible contact forces to judge the stability.
criterion, the other related criteria are unsuitable for judging Poincare return map and capture theory based criteria use the
the stability of rough terrain climbing and do not consider the swing and support legs’ rotational states, which judge stability

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 45

TABLE I
COMPARISON OF RELATED CRITERIA AND OUR CRITERION

by checking whether the robot’s motion states belong to the

prescribed set.
This article’s difficulty lies in how to propose the method of
defining static, quasi-dynamic, and dynamic walking motions
for thruster-assisted legged robots. Moreover, how to judge the
robot’s stability by using the defined motions and considering
the alternating support character is another difficulty.
The main contributions can be specified as follows.
1) We establish a force balancing model of our robot by
considering the fluid forces, hip joint’s driving torque
limit, friction limit, and for rough terrain climbing.
2) We define three motions under the force balancing model,
e.g., static, quasi-dynamic, and dynamic walking, and re-
veal the connection associated with the alternating support
character.
3) We present a stability criterion by using an optimization Fig. 3. Diagram of ith supporting leg that touches rough terrain.
algorithm to find a feasible set of thrusts, which has a
good necessity due to considering the alternating support
character. dynamics caused by ignoring the legs’ masses can make the
4) We propose a stability enhancement method with clear established force balancing model unable to accurately reflect
physical meanings because the desired stability margin the robot’s motion. Therefore, it cannot yield the accurate maxi-
can be achieved only by adjusting the thruster force ac- mum and minimum thruster forces, and the accuracy of stability
cording to the feasible set. judgement can be reduced. However, the unmodeled dynamics
The rest of this article is organized as follows. Section II caused by ignoring legs’ masses are relatively small, and their
establishes a force balancing model for our robot. Section III adverse effects can be ignored.
defines three motions under the established model and presents According to whether touching the rough terrain, the six legs
a model relaxation method to fuse the alternating support charac- can be divided into two categories: supporting legs and swing
ter. Section IV provides a vertical thrust-based stability criterion legs. Fig. 3 shows the diagram of ith supporting leg that touches
and stability enhancement method. Section V presents a 6-D rough terrain. O0 x0 y0 z0 and Oxyz are earth-fixed and body-
thrust vector based stability criterion and stability enhancement fixed frames. Ai and Bi represent the front and end points of
method. Sections VI and VII present simulation and experi- the robot body and ith supporting leg, Ci and Di represent the
mental results, respectively. Finally, Section VIII concludes this hip joint point and touchdown point, r is the radius of the leg,
article. qi represents the rotation angle, where qi = ∠Fi Ci Ti and the
segment Ci Ti is perpendicular to Ci Bi . γi = 2∠Bi Ci Di is the
II. FORCE BALANCING MODEL support angle of the leg. θ is the robot’s pitch angle. fix , fiy , and
fiz are forward, vertical, and lateral touching forces.
A. Framework of the Force Balancing Model Using momentum and angular momentum conservation laws,
Assumption 1: The legs have zero mass, and the robot’s CoM we obtain the framework as

is at the plane’s center point, consisting of six hip joints. Q̇ + w × Q = FB + FG + Fbf + Flf + FT + Af f
Assumption 2: The shape of the leg is a semicircle, and its L̇ + w × L + v × Q = MBG + Mbf + Mlf + MT + Am f
deformation is ignored.
(1)
Assumption 3: Friction coefficients between legs and terrain
are constant and identical. Bt fi  ≤ u (μ)fi ≤ Γi , i = 1, . . . , Ns (2)
Assumption 4: The drive motors of the six legs are identical.
Since the total mass of the six legs accounts for about 3% where Q = m(v + w × rc ) ∈ R3 is the momentum of the robot;
of the robot’s mass, its influence on the robot’s dynamics is L = J0 w + mrc × v ∈ R3 is the angular momentum of the
negligible. In addition, the C-shaped legs are made of hard steel. robot; v ∈ R3 and w ∈ R3 are velocity and angular velocity
Their shape deformations are extremely tiny while impacting of the robot buoyant center; m ∈ R is the mass of the robot,
terrains. Thus, we ignore the legs’ deformations. The unmodeled J0 ∈ R3×3 is the inertial matrix of the robot; rc ∈ R3 is the

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 46 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

where FT = [FT x , FT y , FT z ] , MT = [MT x , MT y , MT z ] ,

T = [T1 , . . . , T8 ] ∈ R8 is the thruster force vector, Ct ∈ R6×8
is the force allocation matrix [4] and it is defined as
⎡ √
2
√
2
√
2
√
2
⎤
0 0 0 0 2 2 2 2
⎢1 1 1 1 0 0√ 0 0√ ⎥
⎢ √ √ ⎥
⎢ 2 2 2 2 ⎥
⎢0 0 0 0 2 − 2 2 − 2 ⎥
⎢ l1 ⎥
⎢ 2 − l21 − l21 l1
0 0 0 0 ⎥
⎢ 2 √ √ √ √ ⎥
⎣0 0 0 0 − 4 2l34 2l34 2l34
− 2l34 ⎦
4 4 4
l2
2
l2
2 − l22 − l22 0 0 0 0
where l1 , l2 , and l3 are three geometrical distances between
thrusters, which are shown in Fig. 4, and l34 = l3 + l4 .
Fig. 4. Placement of thrusters and legs [5]. The robot body’s gravity and buoyancy in the body-fixed
frame can be expressed as
⎡ ⎤ ⎡ ⎤
B sin θ −G sin θ
position vector from the buoyant center to CoM; FB ∈ R3 and FB = ⎣ B cos θ cos φ ⎦ , FG = ⎣−G cos θ cos φ⎦ (4)
FG ∈ R3 are the buoyancy and gravity vectors; MBG ∈ R3 is −B cos θ sin φ G cos θ sin φ
the moment caused by buoyancy and gravity; Fbf ∈ R3 and
where φ is the robot’ roll angle.
Mbf ∈ R3 are the fluid force and fluid moment applied to robot We can deduce the moment vector caused by buoyancy and
body; Flf ∈ R3 and Mlf ∈ R3 are the fluid force and fluid gravity in the body-fixed frame as
moment applied to C-shaped legs; FT ∈ R3 and MT ∈ R3
are the force and moment caused by eight thruster forces; MBG = rc × FG (5)
f = [f1 , . . . , fi , . . . , fN
 
s
] ∈ R3Ns is the set of contact forces
of Ns supporting legs; fi = [fix , fiy , fiz ] ∈ R3 is the contact C. Fluid Forces Applied to Robot Body and Legs
force of ith supporting leg; Af ∈ R3×3Ns and Am ∈ R3×3Ns The fluid forces applied to the robot body are described as
are the contact matrices that sum all contact forces and con-  f
tact moments; Bt = diag(1, 0, 1) ∈ R3×3 is the friction cone Fb = −diag(λ1 , λ2 , λ3 )v̇ − diag(cx , cy , cz )v
(6)
matrix, diag(·) is a function to describe the diagonal matrix; Mbf = −diag(λ4 , λ5 , λ6 )ẇ − diag(mx , my , mz )w
u = [0, μ, 0] ∈ R3 is the matrix representing the contact nor-
mals and the friction coefficient μ;  ·  is the Euclidian norm; where λ1 , λ2 , and λ3 are added masses along the robot body’s
Γi is a positive constant, which is related to the torque limit of longitudinal, vertical, and lateral directions, λ4 , λ5 , and λ6 are
ith supporting leg’s hip joint. added moments of inertia along the three directions. cx , cy , cz
Next, we need to further model, as follows: and mx , my , mz are two groups of hydrodynamic parameters
1) the forces, such as FT , MT , FB , FG , and MBG , caused linearly related to water drags and moments of resistance.
by thrusters, buoyancy, and gravity; Remark 2: Because our robot’s cruising speed is lower than
2) the fluid forces, such as Fbf , Mbf , Flf , and Mlf , related to that of torpedo-like underwater robots, we believe that the
the robot motion states; first-order fluid force accounts for most of the hydrodynamics
3) the contact matrices Af and Am that are associated with and then ignores the second or higher order fluid force. In
robot’s touching state and terrain’s outline; addition, these hydrodynamic parameters such as λ1 , . . . , λ6 ,
4) the positive constant Γi related to ith hip joint’s torque cx , cy , cz , mx , my , and mz are identified from experimental
limit. data by multiple pool cruising tests. In addition, the value of the
Remark 1: The established model in (1) shows that only one angular acceleration ẇ cannot be measured directly by sensors.
touching point is required for each support leg. To ensure this, Here, we use the value of the robot’s current and previous angular
the curvature radius of the terrain near the touching point should velocities and the difference method to obtain the value of ẇ.
be larger than the leg’s radius. In addition, the terrain cannot Before deducing the fluid forces applied to the legs, we need
be sticky because the established model is based on nonnegative to know the velocities of the legs. In addition, we need to lump all
vertical contact forces. Once it is sticky, the vertical contact force fluid forces into the buoyancy centre. Here, we split the deducing
fiy can be less than zero, and the friction constraint in (2) cannot process into three steps: deduce the legs’ velocities, deduce the
be satisfied. fluid forces, and lump the forces into the buoyancy centre.
Step 1: Deduce the legs’ velocities. Based on the placement
of the hip joints in Fig. 4 and the velocity v and angular velocity
B. Thruster Forces, Gravity, and Buoyancy w in the buoyant centre, we can deduce the velocity of kth hip
joint as
According to the thruster placement in Fig. 4, we obtain the
force FT and moment MT caused by eight thrusters as vkh = v + w × rkh , k = 1, . . . , 6 (7)
  where r1h = [d2 , 0, −d1 ] , r2h = [d2 , 0, d1 ] , r3h = [0, 0, −d1 ] ,
FT , MT = Ct · T (3) r4h = [0, 0, d1 ] , r5h = [−d2 , 0, −d1 ] , and r6h = [−d2 , 0, d1 ] ,

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 47

Fig. 5. Velocity of one short arc that belongs to kth leg.

Fig. 6. Contact force of the kth supporting leg.

d1 and d2 are two kinds of distances between hip joints and are Since ∠Ai Ci Bi = θ + qi − π2 , we can derive the position
shown in Fig. 4.
vector from Ci to Di as rit = [licd cos ∠Ai Ci Di , −licd sin ∠Ai
As shown in Fig. 5, we provide the diagram of a short arc and
its velocity viα in kth leg. Based on the geometrical relationship, Ci Di , 0] , where licd = 2r cos( γ2i ), ∠Ai Ci Di = ∠Ai Ci Bi +
2 = θ + qi − 2 + 2 .
γi π γi
we obtain the velocity vkα as
Then, we derive the contact matrices Af and Am as
vkα = vkh + [0, 0, −q̇k ] × rkα (8) 
Af = diag(J1 , J2 , . . . , JNs )
(13)
where rkα = [2r cos αk · cos(θ + qk + αk − π2 ), −2r cos αk · Am = diag(r̂1t J1 , r̂2t J2 , . . . , r̂N
t
J )
s Ns
sin(θ + qk + αk − π2 ), 0] , αk = ∠Bk Ck Ek .
Step 2: Deduce fluid forces. The fluid force applied to the where the hat map (ˆ·) is used to turn a vector into a
short acr can be described as skew-symmetric matrix, Ji denotes a rotation matrix to
project the fi into the body-fixed frame, which is expressed
⎡ ⎤
dFkα = diag(−μl , −μl , 0)vkα · dαk (9) cos(∠Ai Ei Di − π2 ) sin(∠Ai Ei Di − π2 ) 0
where μl is the hydrodynamic coefficient related to the water as Ji = ⎣− sin(∠Ai Ei Di − π2 ) cos(∠Ai Ei Di − π2 ) 0⎦,
density and C-shaped leg’s width. Note that we ignore the fluid 0 0 1
force along the leg’s lateral direction because the C-shaped leg’s where ∠Ai Ei Di = ∠Ai Ci Bi + γi = θ + qi − π2 + γi .
width is much larger than its thickness. According to the geometrical relationship in Fig. 6, we can
The moment on the kth hip joint caused by dFkα can be derive the needed torque of the driving motor in ith hip joint as
deduced as γ  γ 
i i
τi = licd fix cos + licd fiy sin . (14)
dMkα = rkα × dFkα . (10) 2 2
Integrating both sides of (9) and (10), we can deduce the fluid Equation (14) can be rewritten as
 γ 
force and moment applied to the ith hip joint as 2
τi = licd fix 2 + fiy sin
i
+ αi (15)
π 2
Fkα = 0
2
diag(−μl , −μl , 0)vkα · dαk
π (11) where αi = arctan(fix /fiy ).
Mkα = 2
rkα × diag(−μl , −μl , 0)vkα · dαk .
0 Since sin( γ2i + αi ) ≤ 1 and fix ≤ μ|fiy |, we have
Step 3: Lump the forces into the buoyancy center. Lumping 
the forces on six hip joints, we can deduce the force and moment τi ≤ licd 1 + μ2 · u (μ)fi ≤ τmax (16)
applied to the buoyancy center as where τmax is the maximum torque that the driving motor in ith
6 6 hip joint can be provided.
Flf = Fkα , Mlf = (Mkα + rkh × Fkα ). (12) Based on (16), we have u (μ)fi ≤ cd √τmax
2
. Thus, we can
li 1+μ
i=1 i=1 obtain ith supporting leg’s Γi as
τ
D. Contact Matrices and Torque Limit Γi = max . (17)
cd
li 1 + μ2
Fig. 6 shows the contact force fi = [fix , fiy , fiz ] of ith
supporting leg, where i = 1, 2, . . . , Ns , fiy and fix are parallel
III. WALKING MOTION DEFINITION
and perpendicular to the line segment Di Ei , respectively, fiz
is parallel to the axis Oz of the body-fixed frame, Ci and Di We introduce the existing definition of the humanoid robot’s
are points of ith hip joint and touching with the rough terrain, walking motions, which uses the relation between ZMP, CoM,
and Ei is the point that Di projects to the robot body along the and the support area to judge its stability. Inspired by this,
direction of fiy . we propose a new method to define the walking motions for

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 48 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

3) Our definition sorts the walking motions by checking

whether feasible leg-terrain contact forces exist. There-
fore, it is convenient to sort the robot’s motion of climbing
rough terrain. In contrast, the definition in [39] needs to
assume that the terrain is flat. In addition, the use of a
contact force-based force balancing model makes it easy
to fuse other kinds of external forces and thrusts.
We may be unable to find the feasible contact forces (solu-
tions) by using optimization methods when the constraints in
(1) and (2) are too strict. Many factors lead to the constraints’
strictness, e.g., the number of support legs, the amount of inertial,
fluid forces, and thrusts. Next, the existence of feasible solutions
when the robot has 1, 2, and at least three support legs will be
discussed, respectively.
Fig. 7. Diagram of static, quasi-dynamic, and dynamic walking motions. 1) When the robot has only one support leg, the number of
unknown variables in (1) is 3, and there are three contact
forces at the touchdown point. The number of the equality
thruster-assisted legged robots using the established force bal- constraints in (1) is 6. Thus, (1) is overconstrained. That
ancing model to check whether there are feasible contact forces. said, there is no feasible contact force. In such a case, the
Finally, we fuse the alternating support character by relaxing the robot belongs to dynamic walking.
established model as inequality constraints. 2) When the robot has two support legs, the number of
According to the relationship between the CoM, ZMP and the unknown variables in (1) is 6, which is the number of
support convex polygon, Chen et al. [39] define three types of contact forces at the two touchdown points. The number
humanoid robot walking motions. Concretely, when the CoM of inequality and equality constraints are 4 and 6, respec-
and ZMP are located in the support convex polygon area, the tively. Thus, it is highly likely that a feasible solution for
robot belongs to static walking motion. When the CoM is located six variables cannot be found. In such a case, the robot is
in the region of the supporting convex polygon, and the ZMP highly dynamic walking.
is outside the polygon, the robot is in quasi-dynamic walking 3) When the robot has at least Ns (Ns ≥ 3) support legs, the
motion. The robot is in dynamic walking motion when the ZMP number of unknown variables in (1) is 3Ns . The number of
is located outside the support convex polygon. equality constraints in (1) is 6, and the number of inequal-
Before defining the motions, we remove the inertia and fluid ity constraints is 2Ns . Thus, when the inertial forces and
forces from (1) and obtain the static force balancing model as the thrusts are within a reasonable range, feasible solutions
 in 3Ns variables can be found, and the robot may belong
FB + FG + Af f + FT = 0 to static or dynamic walking.
(18)
MBG + Am f + MT = 0. When it belongs to dynamic walking, and its swing leg does
not switch to the support one, the legged robot must fall due to
According to whether there are feasible contact forces f to tip over. However, the legged robot relies on the legs’ alternate
make (1), (2), and (18) satisfy, we can sort the motions into support character to realize the walking. It is still possible to
three categories: static, quasi-dynamic, and dynamic walking make the swing leg switch to the support leg timely by gait
motions, which are shown in Fig. 7. planning, and the robot remains stable.
Definition 1 (Static Walking): If ∃f , satisfying (1), (2), and If the robot’s stability is judged only by the existence of
also ∃f , satisfying (2), (18), the robot is in static walking motion. feasible solutions, the necessity of the stability criterion will
Definition 2 (Quasi-dynamic Walking): If ∃f , satisfying (1) be reduced. Therefore, the equality constraints in (1) should be
and (2), but f satisfies (2) and (18), then the robot is in quasi- appropriately relaxed to improve the necessity of the stability
dynamic walking motion. criterion as follows.
Definition 3 (Dynamic Walking): If f , satisfying (1) and (2), ⎧
then the robot is in dynamic walking motion. ⎪
⎪ −ΩF ≤ −Q̇ − w × Q + (FB + FG + Fbf + Flf
⎨ +Af qf + Af f + FT ) ≤ ΩF
Remark 3: Compared to the definition of the three types of (19)
motions in [39], our definition has the advantage of being close ⎪
⎪ −Ω M ≤ − L̇ − w × L − v × Q + (MBG + Mbf
⎩ f
to the actual situation and can be applied to more universal cases. +Ml + Am qf + Am f + MT ) ≤ ΩM
The advantages lie in the following three aspects.
1) Unlike the definition in [39], in addition to being relevant where ΩF = [ΩF1 , ΩF2 , ΩF3 ] ∈ R3 and ΩM = [ΩM 1
, ΩM2 3 
, ΩM ]
to the gravity and inertia force, our definition is also ∈ R3 are relaxations, ΩF2 = ΩM 2
= 0, ΩF1 , ΩF3 , ΩM1
, and ΩM3

relevant to the fluid and thruster forces. Therefore, our are positive. Note that the values of relaxations reflect the degree
definition is more universal and can be used to define the of losing the model equation in (1) into an inequality in (19), and
motion of thruster-assisted robots and underwater-legged the relaxations contribute to fusing the legged robot’s alternating
robots. support character.
2) Different from the definition in [39], which assumes that The magnitude of the relaxation is related to the leg’s alternate
leg-terrain friction and joints’ driving torques are large touching/swing smoothness and the robot’s mass. When the
enough, our definition relaxes these assumptions. That alternation is smoother, the relaxation should be smaller. In
said, the walking motions sorted by our definition are addition, if the vertical force of the robot body is relaxed, the
closer to the actual situation. supporting leg may leave the ground, and the walking property

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 49

of the legged robot may be lost. Thus, the relaxation amount

ΩF2 is set to zero to maintain the vertical balance. Finally, the
2
magnitude of ΩM represents the degree of losing the yawing
motion, which is unrelated to the alternating support character.
2
Then, we set ΩM = 0.

IV. VERTICAL THRUST-BASED STABILITY CRITERION AND

ENHANCEMENT
We present a criterion to judge the robot’s stability by check-
ing whether the vertical thrust belongs to the range of allowable
vertical forces. On this basis, we propose a method to enhance
the stability by changing the vertical thrust’s value.
When the vertical thrust is too large, the robot may fall due Fig. 8. Motion of rolling on the terrain.
to the support leg’s torque limit. When the vertical thrust is too
small, the forward and lateral contact forces may not be sufficient
to suppress the robot’s inertial and fluid forces, resulting in the
support leg slipping. Setting the appropriate vertical thrusts is Remark 4: We use the optimization method to strip the fea-
necessary so the robot will not fall or slip. We need to find the sible thrusts from multiple variables. Thus, the robot’s stability
maximum/minimum vertical thrusts that can be applied to the can be judged by checking whether the vertical thrust belongs
robot body without losing stability. to the region that consists of maximum and minimum thruster
Specifically, we transform the problem of exploring the max- forces. Unlike the criteria that use multiclass states with different
imum/minimum vertical thrusts, e.g., FTmax min physical meanings, our criterion only relies on one physical
y and FT y , into a
second-order cone programming problem. Then, we use the quantity (vertical thrust). Thus, it has precise physical meanings
sequential least squares programming algorithm to find the and is also convenient for designing the stability enhancement
optimal solutions method.
 Remark 5: Although it possesses the above advantages, the
max FT y or − FT y criterion in this section only uses vertical thrust. Once the robot
(20) cannot maintain the prescriptive balance by adjusting the vertical
s.t. (2) and (19).
thrust, the criterion regards that the robot is unstable. However,
Proposition 1: When the vertical thrust is greater than FTminy
the robot may still be stable when using the 6-D thrust. That said,
or less than FTmax min min the criterion is slightly conservative. Thus, we need to use the
y , i.e., FT y ∈ [FT y , FT y ], the robot is sta-
6-D thrust to improve its necessity and the accuracy of judging
ble, and its stability margin is Ψ = min(FT y − FTmin max
y , FT y − the robot’s stability. In addition, the relaxations in this section are
FT y ). Otherwise, it is unstable. constant and designed artificially. They should be time-varying.
When the robot climbs rough terrains, it will inevitably suffer Thus, we need to derive the expressions of these relaxations.
from various disturbances caused by ocean currents and cables.
The disturbances may make the robot lose its balance. Unlike
V. 6-D THRUST-BASED STABILITY CRITERION AND
the conventional ground legged robots, our robot is actuated
by thrusters and legs simultaneously, where thruster forces can ENHANCEMENT
alter the robot’s dynamics and break or maintain the robot’s This section presents two stability criteria and their corre-
balance. Thus, we present a solution to find the set of feasible sponding stability enhancement method by using the robot’s
thruster forces to keep the robot’s dynamic equilibrium and eight thrusters, which can generate 6-D thruster forces and
prevent the legs’ slippage. moments. First, we provide the method to derive the expressions
The vertical thrust can be set as the average of the obtained of relaxations. Second, we present a method to produce allow-
maximum/minimum vertical thrust to make the robot in the most able thruster forces and moments. Then, we propose a stability
stable state criterion and a method to enhance stability. Finally, we propose
a 6-D thrust based robust stability criterion.
FTmin max
y + FT y
FT y = . (21)
2 A. Derivation of Relaxations
Note that the stability margin of the robot with this verti- This section establishes a 3-D pendulum’s dynamic model
cal thrust is maximum, which can be expressed as Ψmax = to represent the robot’s longitudinal and lateral motions under
FTmax min
y −FT y
2 . In fact, we do not request the robot walk with max- external forces. In addition, we deduce the allowable forces and
imum stability margin because stability and flexibility conflict. moments according to the robot’s allowable swing. Finally, we
Once the robot is in the most stable state, it may be unable to deduce the relaxations’ expressions.
maneuver flexibly. Then, we need to balance the two conflicting As shown in Fig. 8, the pendulum comprises one virtual
requests by setting the stability margin. supporting leg and one robot body, whose endpoint is fixed
Specifically, once the robot is requested to walk with the on the buoyant centre. The legs’ touch angle γ and rotation
desired stability margin Ψd ∈ [0, Ψmax ], we can set the vertical angle q are the average values of all supporting legs, such
 s  Ns
thrust as as γ = N i=1 γi and q = i=1 qi . γ0 and q0 are initial touch
0 0
and rotation angles. (x , y ) and (x, y) are initial and current
FT y = FTmax min
y − Ψd or Ψd − FT y . (22) positions of the pendulum. We also assume that the segment of

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 50 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

D0 D is a straight line, and its slant angle is ϕ. In addition, Fx , Fy , In a word, we need to find the external forces Fx ,
and Fz are forward, vertical, and lateral external forces exerted Fy , Fz , and moments Mx , Mz to make the follow-
on the body. Mx and Mz are roll and pitch external moments, ing conditions limt→t0 +Tf q(t) = qd , limt→t0 +Tf q̇(t) = q̇d ,
respectively. These forces are in the direction of the body-fixed limt→t0 +Tf ζ(t) = ζd , and limt→t0 +Tf ζ̇(t) = ζ̇d hold, where
frame. qd is determined by the gait planner, q̇d is equal to Θmin or Θmax
Actuated by these external forces, the 3-D pendulum’s dy- (Θmin < Θmax ), ζ̇d is equal to Υmin or Υmax (Υmin < Υmax ),
namic model can be established as follows: Θmin , Θmax , Υmin , and Υmax are constants.
⎧
⎨mly ẍ + mlx ÿ = [Fx cos θ + Fy cos(θ + 0.5π)]ly Thanks to the finite-time control theory in [40], we can find
+[Fx sin θ + Fy sin(θ + 0.5π)]lx − Mz (23) the external moments to satisfy the following conditions:
⎩ml2 ζ̈ = F l + M ⎧
⎨u1 = − ac1 V1 1 + bq̇ 2 − a(q − qd − q̈d )
y z y x χ

q̇−q̇d
where ζ is the virtual supporting leg’s dip angle, lx = 2r 2 χ2 (30)
⎩u2 = − c2 mly V2 − mly2 (ζ − ζd − ζ̈d )
cos(0.5γ) sin(π − q − 0.5γ), and ly = 2r cos(0.5γ) cos(π − ζ̇−ζ̇d
q − 0.5γ). 1
Remark 6: Because the allowable variation range of dip angle where χ1 and χ2 are positive constants, 0 < χ1 ≤ 2
1−χ
is slight and often less than 10◦ , the coupling between the and 0 < χ2 ≤ 12 , c1 =
V1 1 (t=t0 )
, V1 =
2
(q−qd ) +(q̇−q̇d ) 2
,
Tf (1−χ1 ) 2
longitudinal and lateral motions is small. Thus, we split the 3-D 1−χ
V2 2 (t=t0 )
locomotion into longitudinal and lateral motions. Moreover, the a = 2mr2 (cos γ + 1), b = mr2 sin γ, c2 = Tf (1−χ2 ) , V2 =
accelerations ẍ and ÿ are coupled, and they are related to the (ζ−ζd )2 +(ζ̇−ζ̇d )2
leg’s rotation angle q and the slant angle ϕ. 2 , q̈d
and ζ̈d are constants, u1 and u2 aim to make
Fig. 8(a) shows that the leg rolls from the initial touch point the equations such as V̇1 = −c1 V1χ1 and V̇2 = −c2 V2χ2 hold.
D0 to the current point D on the slope. Then, we obtain the Because V̇1 = (q̇ − q̇d )(q − qd + q̈ − q̈d ) and V̇2 = (ζ̇ −
position of the point D as ζ̇d )(ζ − ζd + ζ̈ − ζ̈d ), q̇ − q̇d and ζ̇ − ζ̇d will not go to be zero
x = x0D + lD cos ϕ − lx , y = yD
0
+ lD sin ϕ + ly (24) until the V1 and V2 are going to zeros. Thus, q̇ − q̇d and ζ̇ − ζ̇d
will not be zeros when t ∈ (t0 , t0 + Tf ).
where lD is the length of the segment D0 D. Lemma 1: The functions of u1 and u2 in (30) are monotoni-
The local surface that the leg rolls during a complete cycle cally decreasing in the variables of q̇d and ζ̇d , respectively. That
cannot change dramatically, and it is assumed as a slope. Since
said, ∂u 1 ∂u2
∂ q̇d < 0 and ∂ ζ̇d < 0.
the segments O0 D0 and OD are parallel to the segment D0 D,
we have The Lemma’s proof process can be seen in Appendix A.
Substituting the expressions of lx and ly into the expression
q0 − 0.5π + γ0 = q − 0.5π + γ. (25) of the u1 behind (29), we can rewrite u1 as
Thus, we obtain that the length of the segment D0 D as u1 = 2r cos(0.5γ) cos(q + 0.5γ + θ − π)Fx
lD = r(γ0 − γ) = r(q − q0 ). (26) − 2r cos(0.5γ) sin(q + 0.5γ + θ − π)Fy − Mz . (31)
Taking the derivative of both sides of (24) and using the defini- The angle q + 0.5γ + θ would be roughly around π because
tions of lx and ly , we have the robot body is substantially parallel to the segment of
ẋ = r cos ϕq̇ − r cos q q̇, ẏ = r sin ϕq̇ + r sin q q̇. (27) D0 D. Thus, we ignore the rolling dynamics caused by Fy . In
addition, the u1 is equally allocated to Fx and Mz , and the u2
Taking the derivative of (27), we have is equally assigned to Fz and Mx . In such cases, the external
 forces and moments are formulated as
ẍ = r cos ϕq̈ − r cos q q̈ + r sin q q̇ 2 
(28) u1
ÿ = r sin ϕq̈ + r sin q q̈ + r cos q q̇ 2 . Fx = 4r cos(0.5γ) cos(q+0.5γ+θ−π)
(32)
Fy = 0, Fz = 2ly , Mz = − u21 , Mx = u22 .
u2
Substituting (28) into (23) and based on the fact of q + ϕ + γ =
π, we can rewrite the 3-D pendulum model as Since q̇d ∈ [Θmin , Θmax ] and ζ̇d ∈ [Υmin , Υmax ], and
2 2 2
2mr (cos γ + 1)q̈ + mr sin γ q̇ = u1 , mly2 ζ̈ = u2 (29) using the conclusion in Lemma 1, we can derive that the
maximum and minimum of Fx , Fz , Mx , and Mz in (32)
where u1 = (Fx cos θ − Fy sin θ)ly + (Fx sin θ + Fy cos θ)lx as Fxmax = max(Fx (q̇d = Θmin ), Fx (q̇d = Θmin )), Fxmin =
− Mz and u2 = Fz ly + Mx . min(Fx (q̇d = Θmin ), Fx (q̇d = Θmax )), Fzmax = Fz (q̇d =
Next, we will use the 3-D pendulum model, gait planning min min
Θmin ), Fz = Fz (q̇d = Θmax ), Mx = Mx (q̇d = Θmax ),
results, and the allowable lateral motion states to deduce the Mxmax = Mx (q̇d = Θmin ), Mzmin = Mz (q̇d = Θmin ), Mzmax
allowable forces and moments. = Mz (q̇d = Θmax ), where min(·, ·) and max(·, ·) are functions
The gait planner provides all legs’ rotation angles related to to obtain the maximum and minimum values.
time. Then, we define that the virtual leg needs Tf seconds Then, using the maximum and minimum of Fx , Fz , Mx , and
to make the q and q̇ converge to qd and q̇d , where Tf also Mz , we can relax the robot dynamics equations in (1) and convert
means the time spent making the swing legs touching the terrain, them to following inequations:
and qd is the supporting leg’s rotation angle when it begins to ⎧
leave the terrain and to turn into swing leg. In addition, the ⎪
⎪ F ≤ −Q̇ − w × Q + FB + FG + Fbf + Flf
⎨ low +FT + Af f ≤ Fup
allowable lateral motion states are ζd and ζ̇d , which are maximal (33)
dip angle and angular velocity at the moment of t0 + Tf , where ⎪
⎪ M low ≤ − L̇ − w × L − v × Q + MBG + Mbf
⎩ f
t0 is the current moment. +Ml + MT + Am f ≤ Mup

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 51

where Flow = [Fxmin , 0, Fzmin ] , Fup = [Fxmax , 0, Fzmax ] , Nϑ

Mlow = [Mxmin , 0, Mzmin ] , Mup = [Mxmax , 0, Mzmax ] . s.t. ξ = βj ξj∗ ,  β = 1, β 0 (37)

j=1

B. Stability Criterion and Enhancement where ξ is the thrust vector to generate the maximum stability
We choose a set of linearly independent unit vectors to rep- margin, the eight thrusters’ forces can be calculated as T = Ct† ξ ,
resent the directions to find feasible maximum and minimum Ct† is the pseudo-inverse matrix of Ct .
forces. Then, we regard the forces as vertices and construct a In the second method, we also use a quadratic programming
6-D convex set to approximate the real feasible set for reducing algorithm to find the vector ξ that not only inside the convex set
the computational burden. Next, by using the established force but also is closest to previous thrust vector
balancing model, we formulate the problem of finding forces Nϑ
along the designed directions as an optimization problem
min  ξ − ξpre 22
max ϑ
jξ
i=1
Nϑ
s.t. (33), (2) (34)
s.t. ξ = βj ξj∗ ,  β = 1, β 0 (38)
6
where ϑj ∈ R (j = 1, 2, . . . , Nϑ ) is the unit vector to denote j=1

the jth designed direction, ξ = [FT , MT ] ∈ R6 . where ξpre is the thrust vector of the previous moment, which is
After optimizing, we can obtain an extremal vertex called ξj∗ called previous thrust vector, and the eight thrusters’ forces can
along the search direction ϑj . Then, the 6-D feasible region can be described as T = Ct† ξ .
be approximated as a convex set An , which is represented by Remark 7: When the robot is in the most stable situation, it
the vertices representation method as may be unable to maneuver flexibly because stability and flexi-
⎧ ⎫ bility conflict. To balance the two conflicting factors, we present
⎨ Nϑ Nϑ ⎬
the second stability enhancement method, which decreases the
An = βj ξj∗ | β1 ≥ 0, . . . , βNϑ ≥ 0, βj = 1 (35)
⎩ ⎭ variation of thrusts by obtaining the forces closest to the robot’s
j=1 j=1
previous thrusts. Note that the second method can steer the robot
where βj is the jth coefficient. in a critical state once it is unstable in the previous moment.
Proposition 2: The robot is stable if the vertex ξ, composed of
the robot’s thruster force and moment, belongs to An . Otherwise, C. Robust Stability Criterion and Enhancement
it is unstable. The force balancing model in (1) is not accurate enough
Generally, the hyperplane representation method has the ad- because it ignores some dynamics, such as legs’ inertia forces,
vantage of rapidly checking whether a vector is interior to high-order fluid forces, and external disturbances caused by
a polytope by simply checking some given inequations and ocean currents and the electrical cable.
equations. However, it is difficult to represent a convex set if The difficulty of providing an accurate enough robot model
the number of vertices exceeds the vector’s dimension number. lies in the following three aspects.
Thus, we propose a stability checking method based on vertices 1) The coupled dynamics between the legs and robot body
representation. increase the difficulty of modeling. They are challenging
Proposition 3: The vertex ξ belongs to An if there exist co- to decouple for stability analysis.
efficients such as β1 , . . . , βNϑ satisfy the following conditions:
 ϑ ∗
2) Unlike the torpedo-like robots with regular shapes and
1) ξ = N j=1 βj ξj ; high locomotion velocity, our robot has an irregular open-
2) β1 ≥ 0, . . . , βNϑ ≥ 0; frame structure and slow locomotion velocity. It is diffi-
Nϑ
3) j=1 βj = 1.
cult to obtain accurate fluid forces for such a robot and
Here, we use a linear programming algorithm to find the scenario, especially the high-order fluid forces.
feasible coefficients as follows: 3) Ocean current varies in size from place to place, which is
challenging to obtain a priori or measure online.
find β The unmodeled dynamics lead to an inaccurate convex set
that may result in an inaccurate stability assessment. Thus, we
s.t. β = ξ,  β = 1, β 0 (36)
formulate a stricter leg-terrain friction constraint to suppress
where  = [1, 1, . . . , 1] ∈ RNϑ , β = [β1 , . . . , βNϑ ] ∈ RNϑ , uncertainties and present a robust stability criterion that will
 = [ξ1∗ , . . . , ξN
∗
] ∈ R6×Nϑ . result in an accurate stability assessment.
ϑ
Here, we propose two kinds of stability enhancement meth- The force balancing model that includes the unmodeled dy-
ods. The first method aims to make the robot walk with the namics (uncertainties) F ∈ R3 and M ∈ R3 can be estab-
maximum stability margin. The second one finds the stable lished as
⎧ f f
thruster forces that closest approach to the previous thrusts. ⎨Q̇ + w × Q = FB + FG + Fb + Fl + FT + Af f j + F
The first method uses a quadratic programming algorithm to L̇ + w × L + v × Q = MBG + Mb + Mlf + MT + Am f j
f
find the farthest thrust vector ξ away from these vertices in the ⎩
+M,  Bt fij ≤ u fij
convex set
(39)
Nϑ
min  ξ − ξi∗ 22 where F and M are bounded and satisfy that  F < rf ,
i=1  M < rm , rf and rm are two known constants.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 52 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

Theorem 1: If there exists a group of contact forces f = TABLE II

[f1 , . . . , fN
 
s
] ∈ R3Ns to make (40) true, we will find a group GEOMETRIC PARAMETERS OF THE ROBOT

of contact forces f j to make (39) true

⎧
⎪ f
⎨Q̇ + w × Q = FB + FG + Fb + Fl + FT + Af f
f

L̇ + w × L + v × Q = MBG + Mb + Mlf + MT + Am f
f
⎪
⎩
 Bt fi ≤ u fi + li − si
(40) TABLE III
HYDRODYNAMIC PARAMETERS OF THE ROBOT
where li and si are two bounds, si = (rf + rm )σ̄(Eit A† ),
σ̄(·) denotes the largest singular value of the matrix (·),
li = −(rf + rm )  Ein A† , i represents the sequence num-
ber for ith supporting leg, A = [A  
f , Am ] ∈ R
6×3Ns †
, is
the function to calculate the pseudoinverse matrix, Ein =
u , . . . , 01×3 ] ∈ R1×3Ns , Eit = diag(03×3 , . . . ,
[01×3 , . . . , 
ith
Bt , . . . , 03×3 ) ∈ R3×3Ns .
 ith
That said, we can transfer the uncertainties on the robot Compared with the leg-terrain fraction constraint in (40), we
body to the legs’ contacting points and then use stricter friction can obtain a looser friction constraint for ith leg as follows:
constraints to suppress the effects caused by the uncertainties.
This theorem’s proof process is provided in Appendix B. (si − li ) · VI
 Bt fi ≤ u fi + Ns . (42)
Inequation (40) is based on the scaling operations in (47) and
i=1 (si − li )
(48) that aim to find their extreme values. Referring to [21],
we state that these operations need to satisfy the following Based on the established model in (33) and the formulated
conditions. tight friction constraint in (42), we use a quadratic programming
1)  is perpendicular to the vectors of E1n A† , . . . , EN n
s
A† . algorithm to find the feasible force vector ξ that is aligned with
2)  is collinear to the max singular vectors of the predesigned direction ϑj as follows:
E1t A† , . . . , EN
t
s
A† .
We cannot get the extreme values for each leg if these con-
ditions cannot be satisfied simultaneously. Then, we use larger max ϑ
jξ
values to replace the effect caused by the bounded uncertainties s.t. (33), (42), and u fi ≤ Γi . (43)
Fj and Mj . It can reduce the tightness and increase the
conservativeness of the formulated constraint, leading to a more
Then, the 6-D feasible region can be approximated as a convex
conservative stability region.
set An , which is represented as
Next, we will propose a method that finds an integrated
realizable extreme value and then slacks the extreme value for ⎧ ⎫
each leg’s friction constraint. ⎨ Nϑ Nϑ ⎬
Hexapod robots often use a tripod or quadruped gait for loco- An = βj ξj∗ | β1 ≥ 0, . . . , βNϑ ≥ 0, βj = 1 (44)
⎩ ⎭
motion. That is, Ns = 3 or Ns = 4. When Ns = 3, the vector  j=1 j=1
needs to simultaneously satisfy the 2Ns = 6 conditions because
the vectors E1t A† , . . . , E3t A† and the matrices E1n A† , . . . , E3n A† where βj is the jth coefficient, ξj∗ is the extremal vertex along
are different from each other. However, the vector  is 6-D, and the jth search direction.
the maximal number of different conditions that can be satisfied The methods of checking and enhancing the robot’s stability
simultaneously is five rather than six or eight. Thus, the items are the same as the approaches in Section V-B.
are incompatible.
Theorem 2: If there exists a group of contact forces f =
[f1 , . . . , fN
 
s
] ∈ RNs to make (41) true, we will find a group VI. SIMULATION
of contact forces f j to make (39) true We have developed a rough terrain with 6 m in length and
⎧ 2 m in width in the robot simulation environment(Gazebo).
 The
⎪ f f
height function of the terrain is set as y = 8i=0 pi xi , where
⎨Q̇ + w × Q = FB + FG + Fb + Fl + FT + Af f
⎪
p0 = −5, p1 = 0.386, p2 = −2.07, p3 = 3.554, p4 = −2.211,
L̇ + w × L + v × Q = MBG + Mbf + Mlf + MT + Am f
⎪ p5 = 0.624, p6 = 0.081, p7 = 0.00385, p8 = 1.646 × 10−5 .
⎪
⎩ Bt fi ≤ u fi + (sNis−li )·VI We have also developed a virtual robot in Gazebo, which
i=1 (si −li )
(41) shares the same geometric and hydrodynamic parameters as the
real robot. The parameters are provided in Tables II and III.
 s The mass and buoyancy can be changed in Gazebo according
where VI = lI − N i=1 si is an integrated realizable ex-
 s to different simulation objectives. To verify the effectiveness of
treme value, lI = −(rf + rm )  EIn A† , EIn = N i=1 i ∈
E the proposed criterion and stability enhancement method, we let
R1×3Ns . our robot walk on the surface of the rough terrain and carry out
This theorem’s proof process is provided in Appendix C. three groups of simulations.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 53

Fig. 9. Stability probability of robot climbing rough terrain under different relaxations, torque limits, friction coefficients, and accelerations.

A. Stability Analysis With Different Relaxations, Fig. 9(a)– (c) shows stability probabilities under the vertical
Accelerations, Torque Limits, and Friction Coefficients thrust-based stability criterion, where the probability Ps equals
the count of being stable divided by the total count of the
This section aims to observe how different relaxations, accel- climbing process. Fig. 9(a) shows that the greater relaxation
erations, torque limits, and friction coefficient affect the robot’s
results in the larger stability probability. When the acceleration
stability probability. The robot’s mass and buoyancy are set as
reduces, the probability will be increased. Fig. 9(b) demonstrates
15.63 kg and 151.2 N, and the vertical thrust is set as FT y = −40 that a smaller torque limit can result in a smaller stability
N, FT x = FT z = MT x = MT y = MT z = 0. The robot is com-
probability because it will enhance the probability of legs col-
manded to climb the terrain with a tripod gait. Then, we collect
lapsing. Fig. 9(c) shows that the stability probability gets smaller
the corresponding motion data of one climbing process. when the friction coefficient gets smaller. The probability dif-
To verify the effectiveness of the vertical thrust based
ferences caused by friction differences in the first few groups
stability criterion in Section IV, we set nine types of relaxations:
of relaxations are much larger than in the latter groups. The
ΩF 1 = [0, 0, 0] , ΩM 1 = [0, 0, 0] , ΩF 2 = [5, 0, 5] , ΩM 2 = phenomenon arises because the friction coefficient is the domi-
[1, 0, 1] , ΩF 3 = [10, 0, 10] , ΩM 3 = [2, 0, 2] , ΩF 4 =
nant factor in determining the probability when relaxations are
[20, 0, 20] , ΩM 4 = [4, 0, 4] ,

ΩF 5 = [30, 0, 30] , ΩM 5 = small. Fig. 9(d)– (f) provides the stability probabilities under the
[6, 0, 6] , ΩF 6 = [40, 0, 40] , ΩM 6 = [8, 0, 8] , ΩF 7 = 6-D thrust based stability criterion with different accelerations,
[50, 0, 50] , ΩM 7 = [10, 0, 10] , ΩF 8 = [70, 0, 70] , ΩM 8 =
torque limits, and friction coefficients, respectively. The change
[14, 0, 14] , ΩF 9 = [100, 0, 100] , ΩM 9 = [20, 0, 20] . In of the stability probabilities response to accelerations, torque
addition, we decrease the accelerations of the collected data to limits, and friction coefficients are similar to Fig. 9(a)– (c).
its 0.5 and 0.25 to set three kinds of accelerations. That said,
Compared with the relaxations in vertical thrust-based criterion,
they are set as a = a0 , a = 0.5a0 , and a = 0.25a0 , where a0 the deduce of the allowable swing velocity cannot make the
is the original acceleration. We set four types of torque limits, probability drop drastically because it is not the only factor
e.g., τmax = 12 N · m, τmax = 8 N · m, τmax = 4 N · m, and
to determine the Flow , Fup , Mlow , and Mup . The magnitude
τmax = 2 N · m. Finally, the friction coefficients are set as four of stability probabilities in Fig. 9 coincides with the actual
types, e.g., μ = 0.8, μ = 0.4, μ = 0.2, and μ = 0.1. experiences and verifies the two criteria’ effectiveness.
In addition, to verify the effectiveness of the 6-D thrust based
Fig. 10 presents the curves of the maximum vertical thrust
stability criterion in Section V, we set seven groups of allow- FTmin min
able swing velocities Θmin , Θmax and Υmin , Υmax . Specifically, y , minimum thrust FT y , and robot’s thrust FT y under
the seventh relaxation. The trends of these curves contribute
Θmax = −Θmin = Υmax = −Υmin = 9.6, 4.8, 2.4, 1.2, 0.6,
to clarifying how the factors, such as acceleration, maximum
0.3, or 0.01 rad/s. If the allowable swing velocity is set as torque, and friction coefficient, affect the probability. Fig. 10(a)–
9.6 rad/s, we regard this as the seventh group. We regard it as
(c) presents the FTmin max
y , FT y , and FT y with three kinds of
the first group if it is set as 0.01 rad/s. χ1 = 0.5, χ2 = 0.5. The
accelerations. We observe that smaller acceleration can result in
set of accelerations, torque limits, and friction coefficients is the a larger maximum vertical thrust and smaller minimum thrust,
same as the section of verifying the vertical thrust-based stability
which contributes to forming a larger region of allowable thrusts
criterion. We set 76 search directions, which are described as
because it does not need to suppress a force that is too large for
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0 0.4082 −0.4082 inertia under the prescriptive constraints. Fig. 10(c)–(f) provides
⎢0⎥ ⎢ 0 ⎥ ⎢0.4082⎥ ⎢−0.4082⎥ the thrust curves under the torque limits of 8 N · m, 12 N · m, 4
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ N · m, and 2 N · m, respectively. We find that the smaller torque
⎢0⎥ ⎢ 0 ⎥ ⎢0.4082⎥ ⎢−0.4082⎥
⎢0⎥ , · · · , ⎢ 0 ⎥ , ⎢0.4082⎥ , . . . , ⎢−0.4082⎥ . limit can enlarge the value of the minimum thrust and reduce
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣0⎦ ⎣ 0 ⎦ ⎣0.4082⎦ ⎣−0.4082⎦ the probability. Furthermore, Fig. 10(c) and (d) shows that the
0 −1 0.4082 −0.4082 increase of the limit cannot change the value of FTmin y and FT y
min
      significantly because their limits are large enough. Fig. 10(c)
Along the axis ϑ 
1 ,...,ϑ12 Between the axis ϑ 
13 ,...,ϑ76 and (g)–(i) provides the FTmin max
y , FT y , and FT y with four types

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 54 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

Fig. 10. Maximum, minimum, and robot’s vertical thrust of seventh relaxation with different accelerations, maximum torques, and friction coefficients. (a) a =
0.25a0 , τmax = 8 N · m, μ = 0.8. (b) a = 0.5a0 , τmax = 8 N · m, μ = 0.8. (c) a = a0 , τmax = 8 N · m, μ = 0.8. (d) a = a0 , τmax = 12 N · m, μ = 0.8. (e)
a = a0 , τmax = 4 N · m, μ = 0.8. (f) a = a0 , τmax = 2 N · m, μ = 0.8. (g) a = a0 , τmax = 8 N · m, μ = 0.4. (h) a = a0 , τmax = 8 N · m, μ = 0.2. (i) a =
a0 , τmax = 8 N · m, μ = 0.1.

Fig. 11. Extremal thruster forces along the set 76 search directions with three different friction coefficients.

of friction coefficients, e.g., μ = 0.8, μ = 0.4, μ = 0.2, and Figs. 11–13 present one motion state’s extremal vertex (ex-
μ = 0.1. tremal and feasible 6-D thruster force) along the designed
The relaxations and accelerations are related to the smooth- 76 search directions with different friction coefficients, torque
ness of gait planning because nonsmooth gaits can lead to large limits, and upper bounds of unmodeled dynamics. Figs. 11
leg-terrain interaction forces and a poor leg alternating character. and 12 show that a smaller friction coefficient and torque limit
In addition, the friction coefficient is related to the leg-terrain result in a narrower convex set (stable region). That said, it
mechanics. Finally, the torque limit is determined by the chosen raises the difficulty of finding a stable thrust vector. Fig. 13
driving motor in the hip joint. shows that larger upper bounds of unmodeled dynamics can

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 55

Fig. 12. Extremal thruster force along the set 76 search directions with four different torque limits.

Fig. 13. Extremal thruster force along the set 76 search directions with four groups of upper bounds of unmodeled dynamics.

lead to a narrower stable region. Thus, the established force

balancing model should be accurate enough to accurately judge
and effectively enhance the robot’s stability.

B. Stability Analysis With Different Robot Masses

In this section, we conduct two groups of simulations under
four kinds of masses and buoyancies, where the four masses are
set as 67 kg, 37 kg, 15.63 kg, and 5 kg, and the corresponding
buoyancy is 2 N larger than the gravity. The torque limit and
friction coefficient are set as τmax = 2.5 N · m and μ = 0.8. We
use the two groups of simulations to observe the stability proba- Fig. 14. Stability probability under four different masses and relaxations.
bility of the two criteria, and each group has four climbing tests.
We collect the motion states of eight climbing tests and check the
robot’s stability by using the proposed two criteria. In vertical
thrust based criterion, the relaxation is set as {ΩF 5 , ΩM 5 }. In Fig. 14 shows the stability probability of the four tests with dif-
the 6-D thrust-based criterion, the allowable swing velocities are ferent masses. When the mass is larger, the stability probability
selected as Θmax = −Θmin = Υmax = −Υmin = 4.8 rad/s. of the robot is smaller, and the robot can be more unstable. The

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 56 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

TABLE IV
STABILITY PROBABILITIES OF 12 GROUPS

large mass results in a large inertia force and may hinder the robot and 3. The average probabilities of groups 5 and 6 are 79.50%
from suppressing it. The magnitude of stability probabilities in and 85.00%, respectively, which are larger than those of the
Fig. 14 coincides with the robot’s swing and sliding degree, first four groups, and the probability of group 5 is smaller than
which verifies the two criteria’ effectiveness. In addition, our that of group 6. Fig. 15 provides the robot’s vertical thrust and
stability criteria provide a necessary condition that combines the maximum/minimum vertical thrusts of the first test of each
the torque limit and the friction coefficient with the used thrust group. We find that the frequency of vertical thrust belonging to
to design the robot mass for stable locomotion. the region formed by maximum/minimum thrusts in the fifth and
sixth groups is higher than in the first four groups. The mutual
C. Stability Enhancement magnitude character of stability probabilities between the latter
six groups is similar to the first six groups.
To verify the effectiveness of the stability enhancement Based on the abovementioned results, we obtain the following
method, we conduct 12 groups of simulations, where the robot’s findings.
mass and buoyancy are set as 5 kg and 50 N, and each group 1) The average probabilities in group 4 show that the robot
consists of five climbing tests. The vertical thrusts of the groups is probably unstable if the vertical thrust is too large.
1 and 7, 2 and 8, 3 and 9, 4 and 10 are set as −8, −20, −40, and Specifically, the robot may fall because the torque limit
−80N, respectively. The torque limit and friction coefficient are is small, and the hip joint cannot provide enough torque.
set as τmax = 2.5 N · m and μ = 0.8. 2) Compared with the first four groups, the robot in the fifth
We use the vertical thrust-based criterion and stability en- and sixth groups has better stability, which verifies the
hancement method to execute the first six groups of simulations. effectiveness of the vertical thrust-based stability enhance-
Then, we use the 6-D thrust based criterion and method to ment method. In addition, we observe that the stability
perform the latter six groups. In vertical thrust based criterion, worsens if the robot climbs with the maximum stability
we use the fifth relaxation, i.e., {ΩF 5 , ΩM 5 }. The vertical thrusts margin because it will lose part of its maneuverability,
of the latter two groups vary according to the calculated maxi- making it difficult to track the desired motion and reduce
mum and minimum forces. In the 6-D thrust based criterion, the its stability flexibly.
allowable swing velocities are selected as Θmax = −Θmin = 3) The stability probabilities of ten tests in groups 5 and 6 are
Υmax = −Υmin = 4.8 rad/s. Specifically, in the fifth group, the less than 86%, which cannot be improved to 100% because
vertical thrust FT y is designed as FT y = 0.5(FTmin max
y + FT y ), the robot’s stability is enhanced by only relying on the
which aims to force the robot owning largest stability margin. In vertical thrust. Concretely, the large leg-terrain interaction
the sixth group, the vertical thrust FT y is also equal to the one force can result in a large inertial force. Depending on
at the previous moment if the robot is stable. Otherwise, FT y is the vertical thrust may not suppress the force, leading to
designed to make the robot in a critical steady state. That said, it the robot’s instability. Reasonable planning the gait is a
aims to make Ψ = 0 when the robot is unstable. Therefore, the solution to improve the stability probability, but it cannot
vertical thrust is set as be 100%.
 min
FT y , if |FTmin max
y − FT y | < |FT y − FT y |
4) Theoretically, the 6-D thrust-based stability enhance-
FT y = max ment method can ensure the robot’s one hundred percent
FT y , else
stability by taking full advantage of 6-D thruster forces and
where FT y is the vertical thrust of the previous moment. Sim- torques. However, it depends on two preconditions: 1) the
ilarly, the 11th and 12th groups use the stability enhancement established model should be accurate enough; 2) the al-
methods in (37) and (38) to make the robot in a most steady state lowable swing velocities should match the robot’s motion
and critical steady state. planner enough. Once they are inaccurate or match, the
Table IV provides the stability probabilities of 12 groups of 6-D thrust based method will result in inaccurate extremal
simulations. The average probabilities of the first four groups thruster forces. The established model in this article is
are 50.63%, 57.18%, 64.76%, and 37.11%, respectively. The based on four assumptions. Thus, the stability probability
average probability of group 4 is much less than groups 1, 2, in group 12 cannot be 100%.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 57

Fig. 15. Stability probability of six different tests. (a) Test 1 in group 1. (b) Test 1 in group 2. (c) Test 1 in group 3. (d) Test 1 in group 4. (e) Test 1 in group 5.
(f) Test 1 in group 6.

6-D thrust based criteria and stability enhancement methods,

each consisting of five climbing tests. The vertical thrusts of
the groups 1 and 7, 2 and 8, 3 and 9, 4 and 10 are set as −8,
−20, −40, and −80 N, respectively. The method of calculating
the fifth, sixth, 11th, and 12th groups’ thrusts is presented in
Section VI-C. In the experiment of climbing the stairs, the robot
mass, buoyancy, friction coefficient, torque limit, and relaxation
are the same as climbing the curved ramp. We use the method
in group 6 to check whether our stability enhancement method
Fig. 16. Experimental scenario. (a) Curved ramp. (b) Stair. could make the robot climb more challenging rough terrain.

VII. EXPERIMENTS B. Experiment Results

Table V provides the stability probability of six groups of
A. Experiment Setup experiments. The average probabilities of the first four groups
In this section, we experiment to evaluate the proposed sta- are 32.55%, 36.96%, 41.26%, and 28.22%, respectively. The
bility criterion and its enhancement method. We present the average probabilities of groups 1 and 4 are less than those in
experimental scenario in Fig. 16(a), where the robot is on groups 2 and 3. The average probabilities of groups 5 and 6 are
the surface of a man-made rough terrain (curved ramp) and 50.73% and 52.03%, respectively, which are larger than those of
the terrain is on the bottom of a 10 m×10 m×8 m pool. The the first four groups, and the probability of group 5 is smaller than
terrain consists of aluminum alloy skeletons and iron sheets; that of group 6. Fig. 17 presents six snapshots of each group’s
its length and width are 6 m and 2 m, respectively. The height Test 1. The climbing phenomenons of each test are consistent
function of the man-made terrain is similar to the one in the with the probabilities in Table IV. Specifically, the robot in group
simulation, and it can be described as y = 8i=0 pi xi , where one easily slipped and could not complete the climbing task
p0 = −7.47 and the parameters p1 , . . . , p8 are the same ones in because using the small vertical thruster force. The large vertical
simulation. In addition, the robot’s geometric and hydrodynamic force in group 4 can lead to the robot’s shaking, although the
parameters are shown in Tables II and III. As shown in Fig. 16(b), robot can finish climbing. The robot in the other groups can climb
we construct a stair using the aluminium alloy skeletons and the terrain smoothly using a relatively reasonable vertical force.
iron sheets and then present it on the bottom of the pool. We Similar to Section VI-C, the experimental results in groups 7–12
experiment climbing a stair in an indoor pool, where the stair are performed by the 6-D thrust-based stability enhancement
has 12 steps, each with a length 0.15 m, width 1.2 m, and height method. Fig. 18 shows that the vertical thrust-based stability
0.1 m. enhancement method can make the robot climb the stair suc-
In the experiment of climbing the curved ramp, the robot’s cessfully, and its stability probability is 50.27%. We can obtain
mass and buoyancy are 37 kg and 364 N. The friction coefficient the appropriate constant vertical thrust to successfully make
is μ = 0.6. The torque limit is τmax = 16 N · m, and the relax- the robot climb the stairs. However, many trials are needed to
ation is set as ΩF = [50, 0, 50] , ΩM = [10, 0, 10] . Similar find the constant thrust. The stability enhancement method can
to the settings in Section VI-C, we conduct 12 experimental overcome this issue by automatically adjusting based on the
groups to verify the effectiveness of the vertical thrust based and judged stability.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 58 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

TABLE V
STABILITY PROBABILITY OF 12 GROUPS OF EXPERIMENTS

Fig. 17. Snapshot of climbing a rough terrain. (a) Test 1 in group 1, where the robot cannot walk due to legs’ skidding. (b) Test 1 in group 2. (c) Test 1 in group
3. (d) Test 1 in group 4, where the robot cannot complete the climbing task. (e) Test 1 in group 5. (f) Test 1 in group 6. (g) Test 1 in group 11. (h) Test 1 in group 12.

coefficient, and torque limit. Then, we will introduce the two

reasons accounting for the parameters’ difference.
1) In the experiment, the rough terrain’s surface is made of an
iron sheet, which is only three millimetres thick and can
deform slightly under pressure. The buffer caused by the
deformation can reduce the leg-terrain impact forces. The
simulation’s terrain is hard and can result in larger leg-
terrain impact forces. Then, in the simulation, we reduce
the robot’s mass to 5 kg because the smaller mass can
result in smaller impact forces, generating smaller inertia
forces that can be suppressed easily.
Fig. 18. Snapshot of climbing a stair. 2) Since the mass in the simulation is small, we only need
to use much smaller torque limits to steer the legs. Thus,
the torque limit in simulation is set as 2.5 N · m, which is
much smaller than the value of 16 N · m in the experiment.
C. Discussion The application prospects of our methods are list as follows.
We find the results between simulations in Section VI-C and 1) The presented stability criterion and enhancement method
experiments are different, and they arise from the differences can be expanded to other thruster-assisted legged robots
in the chosen parameters, such as the robot’s mass, friction with some modifications. For example, our methods can be

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 59

 
applied to judge and enhance the stability of aerial thruster- = M (2χ1 − 1)(q̇ − q̇d )2 − (q − qd )2 (45)
assisted legged robots, such as the robot in [41]. We need to
2 2 χ1 −1
remove the fluid forces applied to the robot’s body and legs where M = ac1 ( 12 )χ1 · [(q−qd ) (+( q̇−q̇d ) ]
q̇−q̇d )2 > 0.
and modify the contact matrices in Section II-D because its Since χ1 ∈ (0, 0.5], we have 2χi − 1 ≤ 0. Then, we can
legs are not C-shaped. In addition, our criterion can also be
derive that ∂u 1
∂ q̇d < 0. That said, the function u1 in (30) is mono-
used to judge the stability of wheeled robots but not tracked
robots because the wheeled robots touch the terrain with tonically decreasing in the variable of q̇d . Similarly, we have
∂u2
a discrete support manner. We only need to rederive the ∂ ζ̇
< 0.
d
contact matrices according to the specific form-position
relationship between the wheels’ touching points and the APPENDIX B
robot’s CoM in Section II-D. PROOF OF THEOREM 1
2) Our stability criteria can be used as a constraint while
designing the gait planner and controller. In addition, Proof: Based on (1) and (39), we have
we can use our stability enhancement method to obtain Af = Af j +  (46)
the desired stability margin by suppressing the external
disturbances and model uncertainties. where  = [Fj , Mj ] ∈ R6 .
3) Our stability criteria relate to robot mass, torque limits of
Then, we have f j = f − A† . u fi in (1) and u fij in (39)
driving joints, leg-terrain friction coefficient, legs’ shape,
and thruster forces. Then, our criterion can be applied to can be rewritten as Ein f and Ein f j , respectively. Then, we have
guide the design of the robot’s design parameters, such u fij − u fi = Ein (f j − f ) = −Ein A† 
as robot mass and geometric parameters. It can also be
used to guide the choice of joints’ motors and thrusters to ≥ −  Ein A†  ·   ≥ −(rf + rm )  Ein A† ≥ li . (47)
obtain the rational torque limits and thrusts and achieve
the desired locomotion. Similarly, the Bt fi in (1) and Bt fij in (39) can be rewritten as
4) Our stability enhancement method provides a solution to Eit f and Eit f j , respectively. Then, we have
maintain the robot’s stability by regulating the thruster
forces and joint torques separately.  Bt fij = Eit f j = Eit (f − A† ) 
≤ Eit f  +  Eit A†  ≤ Eit f  +σ̄(Eit A† )   
VIII. CONCLUSION
≤ Eit f  +(rf + rm )σ̄(Eit A† ) ≤ Bt fi  +si (48)
In this article, we have presented a new method to define
the three types of walking motion under the established force where σ̄(·) is the largest singular value of the matrix (·).
balancing model. In addition, we have proposed two criteria to Based on the inequations in (47) and (48), we can use
judge the stability of our thruster-assisted underwater hexapod  Bt fi ≤ u fi + li − si to deduce that ∀Fj , Mj , 
robots by checking whether the current thrust belongs to the
allowable region, where the region is calculated by the optimiza- Fj < rf ,  Mj < rm ,  Bt fij ≤ Bt fi  +si ≤
  j
tion method. We have verified our criterion’s effectiveness and u f i + li ≤ u f i .
clarified how factors, such as relaxations, maximum torques,
and friction coefficients, affect the stability through multiple APPENDIX C
comparative simulations and experiments. On this basis, we PROOF OF THEOREM 2
have presented a method to enhance stability by changing the
value of the vertical thrust. Simulation and experiment results Proof: Summing up Ns inequations
 Ns where ith inequation
j
have demonstrated that our method enhances the robot’s stability is shown in (48), we have ( B t fi  −  Bt fi ) ≤
N s i=1
probability by 15%–20%. i=1 si . Referring to the deducing method in (48), we have

Ns
APPENDIX A
PROOF OF LEMMA 1 u (fij − fi ) = EIn (f j − f ) = −EIn A† 
i=1
Taking the partial derivative of the first line of (30) with respect
to q̇d , we have ≥ −(rf + rm )  EIn A†  . (49)
 χ 1 
∂u1 1  s  Ns 
= −ac1 −2χ1 ((q − qd )2 + (q̇ − q̇d )2 )χ1 −1 Thus, we can use N i=1  Bt fi ≤ i=1 u fi + VI to deduce
∂ q̇d 2  Ns
 that ∀ Fj , Mj ,  Fj  < rf ,  Mj  < rm ,
   i=1
((q − qd )2 + (q̇ − q̇d )2 )χ1  Bt fij  ≤ Ns
 B f  + Ns
s ≤ Ns
u 
f + l ≤
+ Ns  j i=1 t i
Ni=1 i i=1 i I
(q̇ − q̇d )2 i=1 u f i , where V I = l I − s
i=1 is . Similarly, the priori
 χ 1 condition to achieve the extreme value VI is that the following
1
= ac1 [(q − qd )2 + (q̇ − q̇d )2 ]χ1 Ns + 1 conditions need to be satisfied simultaneously:
2
! 1)  is perpendicular to the vectors of EIn A† ;
2χ1 1 2)  is collinear to the max singular vectors of
· −
(q − qd )2 + (q̇ − q̇d )2 (q̇ − q̇d )2 E1t A† , . . . , EN
t
s
A† . The hexapod robot in this article walks

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 60 IEEE TRANSACTIONS ON ROBOTICS, VOL. 41, 2025

by a tripod or quadruped gait. That is, Ns + 1 = 4 or 5, [23] L. Chen, R. Cui, W. Yan, and F. Ma, “A stability criterion for hybrid-
which is less than or equal to the maximal provided num- driven underwater bladed legged robot based on capture point theory,”
Acta Automatica Sinica, vol. 50, no. 8, pp. 1565–1576, 2024.
ber 5. Thus, the integrated extreme value VI is realizable. [24] M. Chellapurath et al., “Analysis of station keeping performance of an
underwater legged robot,” IEEE-ASME Trans. Mechatron., vol. 27, no. 5,
pp. 3730–3741, Oct. 2022.
REFERENCES [25] Z. Cai, X. Chen, Q. Li, H. Liu, and Z. Yu, “Design of a felid-like humanoid
foot for stability enhancement,” Biomimetics, vol. 7, 2022, Art. no. 235.
[1] G. Picardi, A. Astolfi, D. Chatzievangelou, J. Aguzzi, and M. Calisti,
[26] H. Liu et al., “Cat-inspired mechanical design of self-adaptive toes for
“Underwater legged robotics: Review and perspectives,” Bioinspiration
a legged robot,” in Proc. IEEE/RSJ Int. Conf. Intell. Robot. Syst., 2016,
Biomimetics, vol. 18, 2023, Art. no. 031001.
pp. 2425–2430.
[2] T. Zhang et al., “From simulation to reality: A learning framework for
[27] J. Zheng et al., “Design, fabrication, and characterization of a hybrid bionic
fish-like robots to perform control tasks,” IEEE Trans. Robot., vol. 38,
spherical robotics with multilegged feedback mechanism,” IEEE Robot.
no. 6, pp. 3861–3878, Dec. 2022.
Autom. Lett., vol. 7, no. 4, pp. 8659–8666, Oct. 2022.
[3] C. Fu and K. Chen, “Section-map stability criterion for biped robots
[28] H. Zhang and A. Song, “Tipover stability enhancement method for a
Part I: Theory,” in Proc. IEEE Int. Conf. Mechatron. Autom., 2007,
tracked mobile manipulator,” in Proc. IEEE Int. Conf. Robot. Autom., 2014,
pp. 1529–1534.
pp. 5158–5164.
[4] F. Ma, W. Yan, L. Chen, and R. Cui, “CPG-based motion planning of
[29] H. Kim, M. Sitti, and T. Seo, “Tail-assisted mobility and stability enhance-
hybrid underwater hexapod robot for wall climbing and transition,” IEEE
ment in yaw and pitch motions of a water-running robot,” IEEE-ASME
Robot. Autom. Lett., vol. 7, no. 4, pp. 12299–12306, Oct. 2022.
Trans. Mechatron., vol. 22, no. 3, pp. 1207–1217, Jun. 2017.
[5] L. Chen, R. Cui, W. Yan, H. Xu, S. Zhang, and H. Yu, “Terrain-adaptive lo-
[30] S. Abeywardena and I. Farkhatdinov, “Towards enhanced stability of
comotion control for an underwater hexapod robot: Sensing leg–terrain in-
human stance with a supernumerary robotic tail,” IEEE Robot. Autom.
teraction with proprioceptive sensors,” IEEE Robot. Autom. Mag., vol. 31,
Lett., vol. 8, no. 9, pp. 5743–5750, Sep. 2023.
no. 1, pp. 41–52, Mar. 2024.
[31] H. Huang, A. Loquercio, A. Kumar, N. Thakkar, K. Goldberg, and J.
[6] M. Calisti, E. Falotico, and C. Laschi, “Hopping on uneven terrains with
Malik, “Manipulator as a tail: Promoting dynamic stability for legged
an underwater one-legged robot,” IEEE Robot. Autom. Lett., vol. 1, no. 1,
locomotion,” in Proc. IEEE Int. Conf. Robot. Autom., 2024, pp. 9712–9719.
pp. 461–468, Jan. 2016.
[32] R. Tong et al., “Design and modeling of an integral molding flexible
[7] G. Picardi, R. Lovecchio, and M. Calisti, “Towards autonomous area in-
tail for robotic fish,” IEEE-ASME Trans. Mechatron., to be published,
spection with a bio-inspired underwater legged robot,” in Proc. IEEE/RSJ
doi: 10.1109/TMECH.2024.3408036.
Int.Conf. Intell. Robot. Syst., 2021, pp. 930–935.
[33] A. W. Winkler, F. Farshidian, D. Pardo, M. Neunert, and J. Buchli, “Fast
[8] C. J. Hu, C. K. Huang, and P. C. Lin, “A torque-actuated dissipative spring
trajectory optimization for legged robots using vertex-based ZMP con-
loaded inverted pendulum model with rolling contact and its use as the
straints,” IEEE Robot. Autom. Lett., vol. 2, no. 4, pp. 2201–2208, Oct. 2017.
template for design and dynamic behavior generation on a hexapod robot,”
[34] Y. De Viragh, M. Bjelonic, C. D. Bellicoso, F. Jenelten, and M. Hutter,
in Proc. IEEE Int. Conf. Robot. Autom., 2015, pp. 5177–5183.
“Trajectory optimization for wheeled-legged quadrupedal robots using
[9] W. C. Lu, M. Y. Yu, and P. C. Lin, “Clock-torqued rolling slip model and
linearized ZMP constraints,” IEEE Robot. Autom. Lett., vol. 4, no. 2,
its application to variable-speed running in a hexapod robot,” IEEE Trans.
pp. 1633–1640, Apr. 2019.
Robot., vol. 34, no. 6, pp. 1643–1650, Dec. 2018.
[35] J. J. Choi, A. Agrawal, K. Sreenath, C. J. Tomlin, and S. Bansal, “Com-
[10] A. W. Winkler, C. D. Bellicoso, M. Hutter, and J. Buchli, “Gait and tra-
putation of regions of attraction for hybrid limit cycles using reachability:
jectory optimization for legged systems through phase-based end-effector
An application to walking robots,” IEEE Robot. Autom. Lett., vol. 7, no. 2,
parameterization,” IEEE Robot. Autom. Lett., vol. 3, no. 3, pp. 1560–1567,
pp. 4504–4511, Apr. 2022.
Jul. 2018.
[36] J. K. Mehr, M. Sharifi, V. K. Mushahwar, and M. Tavakoli, “Intelligent
[11] O. Villarreal, V. Barasuol, P. M. Wensing, D. G. Caldwell, and
locomotion planning with enhanced postural stability for lower-limb
C. Semini, “MPC-based controller with terrain insight for dynamic
exoskeletons,” IEEE Robot. Autom. Lett., vol. 6, no. 4, pp. 7588–7595,
legged locomotion,” in Proc. IEEE Int. Conf. Robot. Autom., 2020,
Oct. 2021.
pp. 2436–2442.
[37] S. Li, Z. Wu, J. Wang, C. Qiu, M. Tan, and J. Yu, “Robust depth
[12] Y. Ding, A. Pandala, C. Li, Y. H. Shin, and H. W. Park, “Representation-free
and heading control system for a novel robotic dolphin with multi-
model predictive control for dynamic motions in quadrupeds,” IEEE Trans.
ple control surfaces,” IEEE Trans. Autom. Sci. Eng., to be published,
Robot., vol. 37, no. 4, pp. 1154–1171, Aug. 2021.
doi: 10.1109/TASE.2024.3380394.
[13] M. Calisti, F. Corucci, A. Arienti, and C. Laschi, “Dynamics of underwater
[38] J. Humphreys, J. Li, Y. Wan, H. Gao, and C. Zhou, “Bio-inspired gait
legged locomotion: Modeling and experiments on an octopus-inspired
transitions for quadruped locomotion,” IEEE Robot. Autom. Lett., vol. 8,
robot,” Bioinspiration Biomimetics, vol. 10, 2015, Art. no. 046012.
no. 10, pp. 6131–6138, Oct. 2023.
[14] M. Vukobratovi and D. Juricic, “Contribution to the synthesis of biped
[39] K. Chen and C. Fu, Humanoid Robot Theory and Technology. Beijing,
gait,” IEEE Trans. Biomed. Eng., vol. BME-16, no. 1, pp. 1–6, Jan. 1968.
China: Tsinghua Univ. Press, 2010.
[15] M. Vukobratovic and B. Borovac, “Zero-moment point—Thirty five years
[40] S. P. Bhat and D. S. Bernstein, “Lyapunov analysis of finite-time differen-
of its life,” Int. J. Humanoid Robot., vol. 1, no. 1, pp. 157–173, 2004.
tial equations,” in Proc. Amer. Control Conf., 1995, vol. 3, pp. 1831–1832.
[16] J. W. Grizzle, G. Abba, and F. Plestan, “Asymptotically stable walking
[41] K. Kim, P. Spieler, E. S. Lupu, A. Ramezani, and S. J. Chung, “A bipedal
for biped robots: Analysis via systems with impulse effects,” IEEE Trans.
walking robot that can fly, slackline, and skateboard,” Sci. Robot., vol. 6,
Autom. Control, vol. 46, no. 1, pp. 51–64, Jan. 2001.
no. 59, 2021, Art. no. eabf8136.
[17] E. R. Westervelt, J. W. Grizzle, and D. E. Koditschek, “Hybrid zero
dynamics of planar biped walkers,” IEEE Trans. Autom. Control, vol. 48,
no. 1, pp. 42–56, Jan. 2003.
[18] H. Hirukawa et al., “A universal stability criterion of the foot contact of
legged robots—adios ZMP,” in Proc. IEEE Int. Conf. Robot. Autom., 2006,
pp. 1976–1983.
Lepeng Chen received the B.Eng. and M.S. degrees
[19] S. Caron, Q. C. Pham, and Y. Nakamura, “ZMP support areas for multi-
in automatic control from Northwestern Polytech-
contact mobility under frictional constraints,” IEEE Trans. Robot., vol. 33,
nical University, Xi’an, China, in 2015 and 2018,
no. 1, pp. 67–80, Feb. 2017.
respectively. He is currently working toward the
[20] T. Bretl and S. Lall, “Testing static equilibrium for legged robots,” IEEE
Ph.D. degree in automatic control with the School
Trans. Robot., vol. 24, no. 4, pp. 794–807, Aug. 2008.
of Marine Science and Technology, Northwestern
[21] H. Audren and A. Kheddar, “3-D robust stability polyhedron in multicon-
Polytechnical University.
tact,” IEEE Trans. Robot., vol. 34, no. 2, pp. 388–403, Apr. 2018.
His research interests are underwater legged
[22] J. Guckenheimer and P. Holmes, Nonlinear Oscillations, Dynamical Sys-
robotics, adaptive control, and motion planning of
tems, and Bifurcations of Vector Fields, vol. 42. Berlin, Germany: Springer,
marine vehicles.
2013.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.
 CHEN et al.: STABILITY CRITERION AND STABILITY ENHANCEMENT FOR A THRUSTER-ASSISTED UNDERWATER HEXAPOD ROBOT 61

Rongxin Cui (Member, IEEE) received the B.Eng. Zhijun Li (Fellow, IEEE) received the Ph.D. degree
degree in autonomic control and the Ph.D. degree in in mechatronics from Shanghai Jiao Tong University,
control science and engineering from Northwestern Shanghai, China, in 2002.
Polytechnical University, Xi’an, China, in 2003 and He is currently a Chair Professor with Tongji Uni-
2008, respectively. versity, Shanghai. He has authored or coauthored
He is currently a Professor with the School of more than 400 papers, where the prestigious contri-
Marine Science and Technology, Northwestern Poly- butions were wearable robotics and bio-mechatronics
technical University. His current research interests systems. His current research interests include wear-
include control and navigation for underwater vehi- able robotics, bio-mechatronics systems, intelligent
cles. control.
Dr. Cui is currently an Associate Editor for IEEE Dr. Li was the recipient of the Distinguished Lec-
TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, and IEEE turer (RAS), the Web of Science Highly Cited Researcher (2019–2023), the 2018
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS. National Tenthousand Talents Program in China, 2016 National Distinguished
Young Scholar (NSFC). He is a Fellow of and AAIA.

Weisheng Yan received the B.Eng. degree in auto-

matic control and the Ph.D. degree in navigation,
guidance, and control from Northwestern Polytech-
nical University, Xi’an, China, in 1991 and 1999, Hui Xu received the Ph.D. degree in mechanical engi-
respectively. neering from Northwestern Polytechnical University,
He is currently a Professor with the School of Xi’an, China, in 2006.
Marine Science and Technology, Northwestern Poly- He is currently an Associate Professor with the
technical University. His current research interests in- School of Marine Science and Technology, North-
clude guidance, navigation, and control of underwater western Polytechnical University. His research inter-
vehicles. ests include underwater robot designing, bioinspired
systems, biomechatronics.

Chenguang Yang (Fellow, IEEE) received the

B.Eng. degree in measurement and control from
Northwestern Polytechnical University, Xian, China,
in 2005 and the Ph.D. degree in control engineering Haitao Yu received the B.Eng. degree in mechan-
from the National University of Singapore, Singa-
ical engineering from Northwestern Polytechnical
pore, in 2010.
University, Xi’an, China, in 2021. He is currently
He is currently the Chair of Robotics with the
working toward the Ph.D. degree in automatic control
Department of Computer Science, University of Liv-
with the School of Marine Science and Technology,
erpool, Liverpool, U.K. His research interest lies in
Northwestern Polytechnical University.
human–robot interaction and intelligent system de-
His research interests include mechanical design-
sign. ing, biomimetic robotics, and mechatronics.
Dr. Yang was a recipient of IEEE TRANSACTIONS ON ROBOTICS Best Paper
Award in 2012 and IEEE Transactions on Neural Networks and Learning
Systems Outstanding Paper Award in 2022, as the lead author.

Authorized licensed use limited to: The Chinese University of Hong Kong CUHK(Shenzhen). Downloaded on July 10,2025 at 14:59:20 UTC from IEEE Xplore. Restrictions apply.