Slide 1

Slide 1 text

Solving Simultaneous Target Assignment and Path Planning Efficiently with Time-Independent Execution Keisuke Okumura & Xavier Defago Tokyo Institute of Technology, Japan ౦ژ޻ۀେֶ 5PLZP*OTUJUVUFPG5FDIOPMPHZ Jun. 21st – 24th, 2022 virtual conf. (Singapore) ICAPS-22

Slide 2

Slide 2 text

/39 2 https://kei18.github.io/tswap pattern formation async execution complete & sub-optimal algorithm (TSWAP) for unlabeled multi-agent pathfinding applicable to both offline & online scenarios ≥1000 agents within 1 sec Summary

Slide 3

Slide 3 text

/39 3 MAPF: Multi-Agent Path Finding given agents (starts) graph goals solution paths without collisions

Slide 4

Slide 4 text

/39 4 Unlabeled/Anonymous MAPF given agents (starts) graph targets solution paths without collisions target assignment

Slide 5

Slide 5 text

/39 5 Motivation optimal sub-optimal (scalable) MAPF unlabeled-MAPF MAPP [Wang&Botea JAIR-11] EECBS [Li+ AAAI-21] PIBT [Okumura+ IJCAI-19] … CBS [Sharon+ AIJ-15] M* [Wagner&Choset AIJ-15] BCP [Lam+ COR-22] … reduction to maximum flow [Yu&LaValle WAFR-13] objective: solve large unlabeled-MAPF with sufficiently good quality in small computation time

Slide 6

Slide 6 text

/39 6 Existing Optimal Algorithm by reducing to maximum flow problem: 𝑂( 𝐴 ⋅ 𝑉 !) * polynomial-time makespan-optimal algorithm exists! [Yu&LaValle WAFR-13] computable but take time (≥1min) fast sub-optimal solver is attractive *assuming |𝐸| = 𝑂(|𝑉|) MAPF benchmarks [Stern+ SOCS-19] 418x530 43,151 257x256 28,178 194x194 14,784 |𝑉|

Slide 7

Slide 7 text

/39 7 optimal linear assignment collision-free scheduling [Yu&LaValle WAFR-13] Hungarian algorithm: 𝑂( 𝐴 ") [Kuhn 1955] lexicographic bottleneck assignment decoupled path planning [Turpin+ AURO-14] [Sokkalingam&Aneja OAL-98] known best algorithm: 𝑂( 𝐴 #) relying on optimal assignments Existing Sub-optimal & Complete Algorithms TSWAP works with arbitrary initial target assignments & applies to online scenarios

Slide 8

Slide 8 text

/39 8 Proposed Algorithm: TSWAP compute arbitrary initial target assignment Step 1. repeat one-timestep path planning with target swapping Step 2. 𝑂( 𝐴 ! ⋅ 𝑑𝑖𝑎𝑚 𝐺 ⋅ (𝛼 + 𝛽)) vertex scoring & deadlock resolution algorithm sketch 2000 agents [Yu&LaValle WAFR-13] optimal algorithm TSWAP 56.3 89.8 230.2 0.4 0.4 1.1 1.36 1.17 1.02 runtime (sec) sub-optimality (makespan) lak303d den520d brc202d evaluation example

Slide 9

Slide 9 text

/39 9 Algorithm Description

Slide 10

Slide 10 text

/39 10 Why Not Arbitrary Assignment + MAPF? => incomplete! need to swap targets on demand

Slide 11

Slide 11 text

/39 11 swap targets rotate targets shortest path move shortest path TSWAP repeat one-timestep path planning with target swapping Step 2. stay currently assigned target stay (otherwise)

Slide 12

Slide 12 text

/39 12 Running Example 1 2 3 planning order: timestep: t

Slide 13

Slide 13 text

/39 13 Running Example move shortest path timestep: t

Slide 14

Slide 14 text

/39 14 Running Example timestep: t

Slide 15

Slide 15 text

/39 15 Running Example swap targets timestep: t

Slide 16

Slide 16 text

/39 16 Running Example timestep: t

Slide 17

Slide 17 text

/39 17 Running Example move shortest path timestep: t

Slide 18

Slide 18 text

/39 18 Running Example timestep: t

Slide 19

Slide 19 text

/39 19 Running Example timestep: t+1

Slide 20

Slide 20 text

/39 20 Running Example timestep: t+1 move shortest path stay

Slide 21

Slide 21 text

/39 21 Running Example timestep: t+1

Slide 22

Slide 22 text

/39 22 Running Example timestep: t+2 fin.

Slide 23

Slide 23 text

/39 23 Completeness the shortest path distance from the current location to the target #targets in the shortest path (excluding endpoints) + Σ agent Proof. Using potential function 0 when all agents are on targets

Slide 24

Slide 24 text

/39 24 the shortest path distance from the current location to the target #targets in the shortest path (excluding endpoints) + Σ agent swap targets rotate targets shortest path stay stay move shortest path

Slide 25

Slide 25 text

/39 25 the shortest path distance from the current location to the target #targets in the shortest path (excluding endpoints) + Σ agent swap targets rotate targets shortest path stay stay move shortest path

Slide 26

Slide 26 text

/39 26 the shortest path distance from the current location to the target #targets in the shortest path (excluding endpoints) + Σ agent swap targets rotate targets shortest path stay stay move shortest path

Slide 27

Slide 27 text

/39 27 the shortest path distance from the current location to the target #targets in the shortest path (excluding endpoints) + Σ agent swap targets rotate targets shortest path stay stay move shortest path

Slide 28

Slide 28 text

/39 28 the shortest path distance from the current location to the target #targets in the shortest path (excluding endpoints) + Σ agent swap targets rotate targets shortest path stay stay move shortest path

Slide 29

Slide 29 text

/39 29 the shortest path distance from the current location to the target #targets in the shortest path (excluding endpoints) + Σ agent swap targets rotate targets shortest path stay stay move shortest path for each timestep, at least one agent performs one of the three actions for each timestep, the potential function decreases

Slide 30

Slide 30 text

/39 30 Initial Target Assignment compute arbitrary initial target assignment Step 1. good assignment? => quick & good quality for certain criteria The paper introduces two algorithms using lazy distance (cost) evaluation resulting in good total/maximum traveling time bottleneck assignment 1. quick with acceptable quality: 𝑂( 𝐴 ⋅ ( 𝑉 + |𝐸|)) greedy assignment with refinement 2.

Slide 31

Slide 31 text

/39 31 Evaluation

Slide 32

Slide 32 text

/39 32 v.s. Makespan-optimal Alg. 0 200 400 600 0 10 20 30 random-64-64-20 64x64 (3,270) random-32-32-20 32x32 (819) makespan runtime (ms) 110 agents TSWAP (bottleneck) TSWAP (greedy) optimal alg. [Yu&LaValle WAFR-13] TSWAP is scalable for graph size solution quality is comparable scalability for graph

Slide 33

Slide 33 text

/39 33 v.s. Makespan-optimal Alg. [Yu&LaValle WAFR-13] TSWAP is scalable for #agents (according to assignment algorithms) 0 500 1000 1500 0 5 10 makespan runtime (ms) TSWAP (bottleneck) TSWAP (greedy) optimal alg. 500 agents 2000 agents (extremely dense) random-64-64-20 64x64 (3,270) scalability for #agents

Slide 34

Slide 34 text

/39 34 Online Planning asynchronous execution

Slide 35

Slide 35 text

/39 35 How to Model Async Execution? time-independent planning [Okumura+ AAAI-21] agents sequentially perform atomic actions we cannot control execution schedule given: start target graph or Online Planning / Policy completeness: regardless of execution schedules, all targets are eventually occupied

Slide 36

Slide 36 text

/39 36 Online TSWAP complete! *this is a centralized algorithm algorithm sketch* compute arbitrary initial target assignment 1. offline phase 2. online phase when is activated: same as offline

Slide 37

Slide 37 text

/39 37 Demo of Time-Independence adaptive to timing uncertainties no synchronous procedure

Slide 38

Slide 38 text

/39 38 8 Robots Demo

Slide 39

Slide 39 text

/39 39 Concluding Remarks TSWAP algorithm our approach target simultaneous target assignment & path planning for indistinguishable agents (unlabeled-MAPF) future directions sub-optimal / complete / quick / scalable applicable to both offline & online decentralization offline planning for async execution c.f., OTIMAPP [Okumura+ IJCAI-22] https://kei18.github.io/tswap