Multi-Robot Motion Planning with Diffusion Models
Authors: Yorai Shaoul, Itamar Mishani, Shivam Vats, Jiaoyang Li, Maxim Likhachev
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We tested MMD s efficacy in learning multi-robot trajectories on increasingly complex maps with varying numbers of holonomic ground robots in a simulated warehouse, modeling robots as 2D disks. Our experimental results from varied motion planning problems in simulated scenarios motivated by logistics environments suggest that our approach scales favorably with both the number of agents and the size of the environment when compared to alternatives. |
| Researcher Affiliation | Academia | 1Carnegie Mellon University 2Brown University EMAIL |
| Pseudocode | Yes | Algorithm 1: MMD-CBS sketch. |
| Open Source Code | Yes | Video demonstrations and code are available at https://multi-robot-diffusion.github.io/. To this end, we make our source code for all MMD algorithms, scripts for data generation, training, and evaluation, and evaluation maps publicly available at https://github.com/yoraish/mmd. |
| Open Datasets | Yes | To this end, we make our source code for all MMD algorithms, scripts for data generation, training, and evaluation, and evaluation maps publicly available at https://github.com/yoraish/mmd. To run our code out of the box, without dataset generation or training, we provide pre-trained models and datasets detailed in Sec. 4 with detailed instructions. |
| Dataset Splits | No | The paper does not provide specific training/validation/test splits for the datasets used to train the diffusion models. It describes how experimental problems are generated and evaluated, but not the data partitioning for model training. For example, it states: "Each experiment with n robots begins by randomly picking start and goal states on a map for various algorithms to compute valid trajectories τ (or MAPF paths Π) between." |
| Hardware Specification | Yes | We implemented all of our algorithms in Python and ran our experiments on a laptop with an an Intel Core i9-12900H CPU, 32GB RAM (5.2GHz), and Nvidia Ge Force RTX 3080Ti Laptop GPU (16 GB). |
| Software Dependencies | No | The paper mentions implementing algorithms in Python and basing implementation on a prior work (Carvalho et al., 2023), but does not provide specific version numbers for Python or any other libraries, frameworks, or tools used. It generally states that exact versions are available in the repository: "We specify the exact versions of libraries and tools required in our repository" (Section 7). |
| Experiment Setup | Yes | The guidance function cost components we used were Jsmooth to encourage dynamically feasible trajectory generation with GPMP, Jobj for obstacle avoidance (both from Carvalho et al. (2023)), and Jc for constraint satisfaction. We set the weights λsmooth = 8e 2, λobj = 2e 2, and λc = 2e 1 for strong soft-constraints and λc = 2e 2 for weak soft-constraints. In our experiments, the size of each local map was 2 2 units, and the diameter of each disk robot was 0.1 units. The radius for CBS sphere constraints was the disk robot radius multiplied by a margin, resulting in a radius value of 0.12 units, and the time interval t was 0.08 seconds (2 time steps). |