Research map / Image generation / Autoregressive and latent generation
Multi-Robot Motion Planning with Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MMD, which composes single-robot diffusion motion planners with constraint-based MAPF search (PP, CBS, ECBS variants) via spatio-temporal sphere guidance, and sequences local diffusion models for large maps.
Problem
Multi-robot motion planning that produces collision-free trajectories matching demonstration data; multi-robot diffusion models suffer high sample complexity and fail to scale beyond a few robots.
Approach
Learn only single-robot diffusion models and impose collision-avoidance as soft sphere constraints in the denoising guidance, with constraint placement chosen by MAPF search strategies (prioritized planning, CBS, ECBS, plus warm-started replanning); long-horizon plans are built by sequencing local diffusion models via inpainting-style connection constraints.
Key results
- MMD-xECBS solved all 300 tests with 3, 6, and 9 robots on Empty/Highways where MPD-Composite achieved 0% success with 9 robots.
- On Highways with 9 robots, MMD-xECBS achieved 100% success with data adherence 0.97 versus A*Data-ECBS 0.88 (no data-driven cost at equal quality), and MMD-xECBS scaled to 40 agents in additional tests.
- MMD-PP achieved lower success than CBS-based variants in congested maps (e.g., 30% on Conveyor with 9 robots), while CBS-based methods failed mainly by exceeding the 60-second time limit.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MMD-xECBS instead of MPD-Composite | Multi-robot trajectory generation · Empty, 6 robots | success rate: 62% -> 100% | Helped |
| Use MMD-xECBS instead of MPD-Composite | Multi-robot trajectory generation · Empty, 9 robots | success rate: 0% -> 100% | Helped |
| Use MMD-xECBS instead of MPD-Composite | Multi-robot trajectory generation · Highways, 6 robots | success rate: 0% -> 98% | Helped |
| Use MMD-xECBS instead of MPD-Composite | Multi-robot trajectory generation · Highways, 9 robots | success rate: 0% -> 96% | Helped |
| Use statistical data cost maps with A*Data-ECBS rather than uniform-cost A*-ECBS | MAPF planning · Highways | data adherence: +35% | Helped |
| Use statistical data cost maps with A*Data-ECBS rather than uniform-cost A*-ECBS | MAPF planning · Highways | success rate | Hurt |
| Add previous-trajectory reuse during CBS replanning (ablation) | MMD-CBS and MMD-ECBS · Multi-robot planning | performance | Helped |
| Use MMD-PP rather than CBS-based MMD strategies | MMD · Congested maps | success rate | Hurt |
| Use MMD-xECBS rather than other MMD algorithms | MMD · Multi-robot planning | success rate | Helped |
| Sequence local diffusion models for long-horizon planning | MMD · Larger maps and long planning horizons | data adherence | Helped |
| Train a joint multi-robot composite model instead of using single-robot models with MMD | Multi-robot trajectory generation · Empty and Highways, increasing robot counts | success rate: 100% -> 62% -> 0% | Hurt |
| Use MMD-xECBS instead of A*Data-ECBS | Multi-robot planning · Conveyor, 3 robots | data adherence: 0.03 -> 0.87 | Helped |
Surprising findings
- Composite multi-robot diffusion model with 6 robots achieved only 62% success on Empty and 0% on Highways, while the per-robot MMD approach stayed near 100%.
- CBS-based MMD methods struggled on the circle scenario, while the soft-constraint PP variant navigated congestion better in that setup.
Limitations
- CBS-based MMD variants are slow due to sequential high-level search and replanning; runtime grows with the number of robots
- MMD-PP soft constraints do not guarantee collision-free solutions, causing failures in congested maps
- Centralized full-horizon planning assumes known obstacle motion and perfect execution
Experimental setup
Datasets: simulated single-robot motion demonstrations on Empty, Highways, Conveyor, Drop-Region maps multi-robot trajectory data for MPD-Composite baseline on Empty and Highways
Benchmarks: Empty map Highways map Conveyor map Drop-Region map circle and weave free-space scenarios large tiled maps (2x2, 3x3)
Metrics: success rate data adherence planning time average acceleration
Baselines: MPD-Composite A*Data-ECBS A*-ECBS MMD-PP MMD-CBS MMD-ECBS
Builds on: Diffuser MPD (motion planning diffusion) Conflict-Based Search ECBS Prioritized Planning
Closest papers by meaning
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.92)
- Simultaneous Multi-Robot Motion Planning with Projected Diffusion Models (ICML 2025 · similarity 0.91)
- Simulation-Informed Diffusion for Decentralized Multi-robot Motion Planning (NeurIPS 2026 · similarity 0.91)
- Generative Trajectory Stitching through Diffusion Composition (NeurIPS 2025 · similarity 0.89)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.88)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.