Accelerating Search-Based Planning for Multi-Robot Manipulation by Leveraging Online-Generated Experiences

TL;DR

This paper addresses the challenge of planning for multiple robotic arms operating in a shared workspace, where the high-dimensional composite configuration space makes traditional planning methods intractable. It introduces an experience-acceleration framework that reuses online-generated planning experiences inside Conflict-Based Search-based MAPF solvers, instantiated as xCBS and xECBS, to speed up replanning while preserving completeness and bounded sub-optimality. Theoretical analysis shows that the low-level xWA* retains completeness and a sub-optimality bound $w^L$, and the high-level CBS search maintains a bound $w^H$, yielding an overall $w^H w^L C^*$-optimality; ECBS-style bounding is preserved in the accelerated variants. Empirically, the approach scales to up to 10 arms, outperforms baselines in planning time and success rate—especially in cluttered, closely interacting scenarios—highlighting xECBS as particularly effective for practical multi-arm manipulation tasks.

Abstract

An exciting frontier in robotic manipulation is the use of multiple arms at once. However, planning concurrent motions is a challenging task using current methods. The high-dimensional composite state space renders many well-known motion planning algorithms intractable. Recently, Multi-Agent Path-Finding (MAPF) algorithms have shown promise in discrete 2D domains, providing rigorous guarantees. However, widely used conflict-based methods in MAPF assume an efficient single-agent motion planner. This poses challenges in adapting them to manipulation cases where this assumption does not hold, due to the high dimensionality of configuration spaces and the computational bottlenecks associated with collision checking. To this end, we propose an approach for accelerating conflict-based search algorithms by leveraging their repetitive and incremental nature -- making them tractable for use in complex scenarios involving multi-arm coordination in obstacle-laden environments. We show that our method preserves completeness and bounded sub-optimality guarantees, and demonstrate its practical efficacy through a set of experiments with up to 10 robotic arms.

Figures (4)

• Figure 1: Eight robotic manipulators, each of 7-DoF, collaborating in a shelf-rearrangement pick and place task. Planning concurrent motions for all arms requires a motion planner capable of efficiently exploring a single arm's high-dimensional state space and reasoning about the motions of multiple robots operating in the shared task space.
• Figure 2: An illustration of our proposed algorithm accelerating a single agent search on a four-connected grid via reusing previous search efforts. (a) A single-agent path from $S$ to $G$ computed in a previous iteration. (b) Upon imposing a new constraint on the agent, shown in red, replanning is required. The previous path is drawn in light gray. (c) Upon expansion of node $S$, a prefix $\{A, B, C\}$ of the experience path is added to OPEN along with all other successors of $S$. (d) shows two steps: node $C$ is selected for expansion from OPEN, and in the next iteration node $J$ is expanded from OPEN. Upon expanding $J$, a segment of the experience is added to OPEN, since one of $J$'s successors is equivalent to a node in the experience. (e) Finally, $G$ is expanded from OPEN and the search terminates and recovers a path. In this example, the work done by xWA* (Alg. \ref{['alg:low-level']}) is smaller than that of its previous iteration. By reusing experience, the intermediate nodes expanded are $C$ and $J$.
• Figure 3: Evaluating the real-world applicability of planning algorithms. Left: evaluation scenes, with 8-arm shelf rearrangement and 4-arm bin-picking. Middle: Comparing planning time and cost for each row-name planner relative to the column-name planner. The values offer a fair comparison by considering only successful runs in both planners. For instance, xECBS has shorter planning times (red, above 100%) and lower solution costs comparable to other CBS-based approaches (blue, around or below 100%). xECBS is faster and finds short paths. Right: Success rate and mean$\pm$standard deviation among successful runs.
• Figure 4: Scalability analysis. Top: our test scenes with 2, 4, 6, 8, and 10 robots. Bottom, from left to right: (a) success rate of methods in the $50$ planning problems of each scene. xECBS scales better than competing methods. (b) Average planning time in successful runs. xECBS maintains a lower planning time as the number of robots increases. (c) Cost per agent in successful runs. All CBS-based methods maintain similar costs while PP and sampling-based methods eventually produce worse paths. (d) Average number of collision checks. (b,c,d) report on planners with at least $5\%$ success to avoid unrepresentative data.

Theorems & Definitions (5)

• Lemma 1
• Lemma 2
• Theorem 1
• Lemma 3
• Theorem 2