Multi-Agent Path Finding (MAPF)
• Multi-robot path finding
• Given: a number of robots (each with a start and goal
location) and a known environment
• Task: find collision-free paths for the robots from their
start to their goal locations that minimize some objective
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
• Optimization problem with the objective to minimize
task-completion time (called makespan) or
the sum of travel times (called flowtime)
4-neighbor grid
Multi-Agent Path Finding (MAPF)
• Application: Amazon fulfillment centers
4-neighbor grid [work by Kiva Systems/Amazon Robotics, not me]
Multi-Agent Path Finding (MAPF)
• Optimal MAPF algorithms
– Theorem [Yu and LaValle]: MAPF is NP-hard to solve
optimally for makespan or flowtime minimization
[www.random-ideas.net]
• Bounded-suboptimal MAPF algorithms
– Theorem: MAPF is NP-hard to approximate within any
factor less than 4/3 for makespan minimization on
graphs in general
Multi-Agent Path Finding (MAPF)
A B C D E
1 S2
2 S1
3 G1
4 G2
4-neighbor grid
A B C D E
1 S2
A*-Based Search 2
3
S1
G1
4 G2
• A*-based search: Optimal (or bounded-suboptimal) MAPF solver
A2
B1
A2 A2 A3
B1 C1 … B2
A B C D E
1 S2
Conflict-Based Search 2
3
S1
G1
4 G2
• Conflict-based search [Sharon, Stern, Felner and Sturtevant]:
Optimal (or bounded-suboptimal) MAPF solver that plans for
each agent independently
Add constraint: Add constraint:
the red agent is not allowed the blue agent is not allowed
to be in cell D3 at time 4 to be in cell D3 at time 4
4-neighbor grid [work by Ben-Gurion University of the Negev, not me]
A B C D E
1 S2
Conflict-Based Search 2
3
S1
G1
4 G2
• Conflict-based search [Sharon, Stern, Felner and Sturtevant]:
Optimal (or bounded-suboptimal) MAPF solver that plans for
each agent independently
Add constraint: Add constraint:
the red agent is not allowed the blue agent is not allowed
to be in cell D3 at time 4 to be in cell D3 at time 4
To minimize the sum of the travel times of all agents
perform a best-first search on this tree with
• g = sum of travel times of all agents
4-neighbor grid • h=0
A B C D E
1 S2
Improvement 1 2
3
S1
G1
4 G2
• Use more informed (= non-zero) h-values
Add constraint: Add constraint:
the red agent is not allowed the blue agent is not allowed
to be in cell D3 at time 4 to be in cell D3 at time 4
The sum of travel times of any collision-free
solution is at least 11.
4-neighbor grid
A B C D E
1 S2
Improvement 1 2
3
S1
G1
4 G2
• Use more informed (= non-zero) h-values
h=1
cardinal conflict
The sum of travel times of any collision-free
solution is at least 11.
4-neighbor grid
Improvement 1
• Use more informed (= non-zero) h-values
minimum vertex
cover is 3
h=3
4-neighbor grid
A B C D E
1 S2
Improvement 2 2
3
S1
G1
4 G2
• Symmetry breaking of rectangle conflicts
Add constraint: Add constraint:
the red agent is not allowed the blue agent is not allowed
to be in cell D3 at time 4 to be in cell D3 at time 4
The sum of travel times of any collision-free
solution is at least 11 but conflict-based search
does not detect it right away.
4-neighbor grid
A B C D E
1 S2
Improvement 2 2
3
S1
G1
4 G2
• Symmetry breaking of rectangle conflicts
Add constraint: Add constraint:
the red agent is not allowed the blue agent is not allowed
to be in cell D3 at time 4 to be in cell D3 at time 4
A B C D E A B C D E
1 S2 1 S2
2 S1 2 S1
3 G1 3 G1
4 G2 4 G2
A B C D E
1 S2
Improvement 2 2
3
S1
G1
4 G2
• Symmetry breaking of rectangle conflicts
4-neighbor grid
A B C D E
1 S2
Improvement 2 2
3
S1
G1
4 G2
• Symmetry breaking of rectangle conflicts
Add constraint: Add constraint:
the red agent is not allowed the blue agent is not allowed
to be in cell D3 at time 4 to be in cell D3 at time 4,
or in cell D2 at time 3 in cell C3 at time 3
or in cell B2 at time 2
barrier constraints
A B C D E
1 S2
Improvement 3 2
3
S1
G1
4 G2
• Disjoint splitting
Add constraint: Add constraint:
the red agent is not allowed the blue agent is not allowed
to be in cell D3 at time 4 to be in cell D3 at time 4
A B C D E A B C D E
1 S2 1 S2
2 S1 2 S1
3 G1 3 G1
4 G2 4 G2
A B C D E
1 S2
Improvement 3 2
3
S1
G1
4 G2
• Disjoint splitting
Add constraint: Add constraint:
the red agent is not allowed the red agent must
to be in cell D3 at time 4 be in cell D3 at time 4
A B C D E A B C D E
1 S2 1 S2
2 S1 2 S1
3 G1 3 G1
4 G2 4 G2
A B C D E
1 S2
Improvement 3 2
3
S1
G1
4 G2
• Disjoint splitting
Add constraint: Add constraint:
the red agent is not allowed the red agent must
to be in cell D3 at time 4 be in cell D3 at time 4
which implies that all
the other agents are not
allowed to be in cell D3
at time 4
4-neighbor grid
Improvement 3
• Disjoint splitting
A B C D E
1 S2
2 S1 G1
3 G2
Original CBS Disjoint CBS
Pruned
4-neighbor grid
Improvement 4
• Rapid random restarts help to solve
more multi-agent path finding
problems within a given runtime limit.
• Here: We randomize the ordering in
which the agents plan their paths in
the high-level root node.
runtime
runs time limit 38 “easy” 12 “hard” 50 total
1 300 sec 100.00% 0.00% 76.00%
3 100 sec 97.65% 96.87% 97.60%
5 60 sec 98.57% 98.81% 98.70%
Conflict-Based Search
Optimal Bounded-Suboptimal
MAPF Planning MAPF Planning
CBS with heuristics
and symmetry breaking (E)CBS with highways
and parallel runs
CBS with heuristics
(E)CBS
CBS
CBS (E)CBS
CBS with heuristics
CBS with heuristics (E)CBS with highways
and symmetry breaking and parallel runs
4-neighbor grid
Conflict-Based Search
[Wurman, D’Andrea and Mountz]
4-neighbor grid
Lifelong Multi-Agent Path Finding
• Runtime on 135x31 grids
• 250 agents and 20,000 random pickup-and-delivery tasks
• Makespan ≈ 0.5 hour
• Mean total planning time ≈ 10s
4-neighbor grid
More Information on MAPF
• Go to mapf.info for more information on MAPF
Acknowledgments
• This tutorial reported on joint work with a large
number of collaborators (including students) from the
University of Southern California and elsewhere. We
would like to acknowledge their contributions
• Special thanks to K. Arras, A. Arunasalam, N. Ayanian, E.
Boyarski, T. Cai, D. Chan, H. Choset, L. Cohen, K. Daniel, A.
Felner, D. Harabor, C. Hernandez, W. Hoenig, S. Jahangiri, T. K.
S. Kumar, M. Likhachev, H. Ma, P. Meseguer, A. Nash, L.
Palmieri, G. Sharon, X. Sun, P. Stuckey, N. Sturtevant, C. Tovey,
T. Uras, G. Wagner, H. Xu, W. Yeoh, S. Young and D. Zhang
• Thanks to Amazon, ARO, IBM, JPL, NSF, ONR for
funding!
Acknowledgments
• Please visit idm-lab.org/projects.html for more
information, pointers to the literature and our
publications
• If you have any interesting ideas, please send me an
email:
[email protected]
