Multi-Robot Motion Planning with Diffusion Models

TL;DR

MRMP is computationally hard and data-hungry. The paper proposes MMD, a framework that uses single-robot diffusion priors guided by MAPF-inspired constraints to generate collision-free, data-aligned trajectories for many robots without requiring multi-robot demonstrations. It introduces five constraint-based variants and a sequencing approach to enable long-horizon planning, demonstrating scalability to dozens of robots in logistics-like environments and outperforming composite diffusion baselines and MAPF with learned costs in data adherence and success rates. The work includes reproducible resources, including code, datasets, and pretrained models, to enable broader adoption and extension.

Abstract

Diffusion models have recently been successfully applied to a wide range of robotics applications for learning complex multi-modal behaviors from data. However, prior works have mostly been confined to single-robot and small-scale environments due to the high sample complexity of learning multi-robot diffusion models. In this paper, we propose a method for generating collision-free multi-robot trajectories that conform to underlying data distributions while using only single-robot data. Our algorithm, Multi-robot Multi-model planning Diffusion (MMD), does so by combining learned diffusion models with classical search-based techniques -- generating data-driven motions under collision constraints. Scaling further, we show how to compose multiple diffusion models to plan in large environments where a single diffusion model fails to generalize well. We demonstrate the effectiveness of our approach in planning for dozens of robots in a variety of simulated scenarios motivated by logistics environments. View video demonstrations and code at: https://multi-robot-diffusion.github.io/.

Figures (8)

• Figure 1: MMD-CBS sketch. Colored lines are only in MMD-PP, MMD-ECBS
• Figure 2: A comparison between MMD and "composite" diffusion models that generate trajectories for all agents at once. We observed consistent performance from MMD but a sharp decrease for the baseline, unable to produce valid solutions for $9$ agents (denoted as zero adherence score). Since MMD uses the same single-agent model for all robots in these experiments, it is straightforward to keep increasing the number of agents without needing any retraining or new datasets.
• Figure 3: Analysis of success rates and data adherence scores, in randomly generated planning queries, of all MMD instantiations and a MAPF method with and without a learned cost map. The left column shows our test maps, the middle column compares success rates across $10$ trials per robot count, and the right column presents the average data adherence scores.
• Figure 4: Scalability tests in high-congestion free-space MRMP. Circle (top row) asks robots to swap positions between opposite points on the perimeter. Weave (below), asks robots to exchange positions along uniformly spaced boundary points. Length is zero for failed problems (MMD-PP failures were due to yielding invalid solutions, and other methods failed by exceeding 240 seconds).
• Figure 5: Experimental setup and results for scaling MMD to larger environments and longer planning horizons. MMD still relies on single agent data in small, local maps and does not require training of new networks for this scale-up. We evaluate three MMD variants across two large maps made of tiled local maps to cover a significantly larger area.