Multi-Robot Motion Planning with Diffusion Models
Yorai Shaoul, Itamar Mishani, Shivam Vats, Jiaoyang Li, Maxim, Likhachev
TL;DR
This paper introduces MMD, a novel method combining diffusion models and classical search to generate collision-free multi-robot trajectories from single-robot data, enabling scalable multi-robot planning in large environments.
Contribution
The paper presents a new approach that integrates diffusion models with classical planning to handle multi-robot scenarios using only single-robot data, addressing scalability and generalization.
Findings
Effective planning for dozens of robots in simulations
Generates collision-free trajectories conforming to data distributions
Scales to large environments with multiple diffusion models
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…
Peer Reviews
Decision·ICLR 2025 Spotlight
1. This paper provides a very thorough discussion of how to combine MAPF algorithms with diffusion models and demonstrate its ability to scale to multi-agent settings. 2. The experiment setting is diverse with varying levels of difficulties. 3. The writing is clear and structured and the paper is easy to read.
1. It would be good to discuss other search algorithms in multi-agent motion planning or at least discuss why only A* is chosen 2. In Sec. 4.2, it would be good to discuss the computational complexity between A* and MMD. 3. The explanation of the motion pattern is insufficient. Then it becomes a bit confusing in Sec 3.3 how this task-relevant local diffusion model works and also in Sec. 4 what the data adherence means.
* The questions of multi-robot motion planning that the paper aims to solve is meaningful and it has many real-world applications like automated warehouses. * The method proposed by the paper is diretly and simple. They leverage state-of-art single-robot diffusion planning methods with classifical MAPF algorithms to solve multi-robot motion planning problems. * They also evaluate their approaches in a variety of simulated scenarios and show better performance. They also conduct rich ablation stu
* The novelty of the method seems limited. The paper seems simply combine diffusion planning methods and classifical MAPF algorithms. In the method section (3.2), It's hard to judge which parts are from classifical MAPF algorithms, which parts are contributed by the paper to adapt MAPF algorithms into diffusion planning methods. * Lack of some experimental details and results. See questions. * The methods are not very efficient when the number of agents increases.
This paper introduces a novel approach for multi-robot motion planning, particularly in combination MAPF with diffusion models. This approach sidesteps the challenge of multi-agent data collection by training on single-robot data, allowing the method to efficiently address the curse of dimensionality typically seen in large-scale multi-agent systems. Also, it incorporates MAPF-based constraint methods, such as Conflict-Based Search, to dynamically enforce collision-free paths. By using an ensem
A major limitation is the reliance on single-robot data for policy learning. This assumption may not be sufficient for complex, dynamic environments where multi-robot interactions are essential. While multiple experiments demonstrate the effectiveness of this approach, exploring the differences between multi-agent and single-agent data, and how single-agent data can benefit cooperative multi-robot motion planning, would add valuable insight.
Code & Models
Videos
Taxonomy
TopicsRobotic Path Planning Algorithms · DNA and Biological Computing · Modular Robots and Swarm Intelligence
MethodsDiffusion