Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Decoupled MPPI-Based Multi-Arm Motion Planning

Dan Evron, Elias Goldsztejn, Ronen I. Brafman

TL;DR

MR-STORM extends GPU-accelerated MPPI-based STORM to decentralized multi-arm motion planning in dynamic environments by incorporating a dynamic obstacle cost and a distance-based prioritization scheme. Each arm computes its own plan while sharing near-term trajectories, which are treated as dynamic obstacles by others, enabling reactive, per-arm coordination without a central supervisor. The approach shows clear empirical advantages over state-of-the-art centralized and decentralized baselines across 120 simulated scenarios in Isaac-Sim, with negligible runtime overhead relative to control frequency. This yields a practical, scalable framework for real-time multi-arm manipulation in cluttered and uncertain environments, preserving STORM's performance benefits while addressing inter-arm interactions.

Abstract

Recent advances in sampling-based motion planning algorithms for high DOF arms leverage GPUs to provide SOTA performance. These algorithms can be used to control multiple arms jointly, but this approach scales poorly. To address this, we extend STORM, a sampling-based model-predictive-control (MPC) motion planning algorithm, to handle multiple robots in a distributed fashion. First, we modify STORM to handle dynamic obstacles. Then, we let each arm compute its own motion plan prefix, which it shares with the other arms, which treat it as a dynamic obstacle. Finally, we add a dynamic priority scheme. The new algorithm, MR-STORM, demonstrates clear empirical advantages over SOTA algorithms when operating with both static and dynamic obstacles.

Decoupled MPPI-Based Multi-Arm Motion Planning

TL;DR

MR-STORM extends GPU-accelerated MPPI-based STORM to decentralized multi-arm motion planning in dynamic environments by incorporating a dynamic obstacle cost and a distance-based prioritization scheme. Each arm computes its own plan while sharing near-term trajectories, which are treated as dynamic obstacles by others, enabling reactive, per-arm coordination without a central supervisor. The approach shows clear empirical advantages over state-of-the-art centralized and decentralized baselines across 120 simulated scenarios in Isaac-Sim, with negligible runtime overhead relative to control frequency. This yields a practical, scalable framework for real-time multi-arm manipulation in cluttered and uncertain environments, preserving STORM's performance benefits while addressing inter-arm interactions.

Abstract

Recent advances in sampling-based motion planning algorithms for high DOF arms leverage GPUs to provide SOTA performance. These algorithms can be used to control multiple arms jointly, but this approach scales poorly. To address this, we extend STORM, a sampling-based model-predictive-control (MPC) motion planning algorithm, to handle multiple robots in a distributed fashion. First, we modify STORM to handle dynamic obstacles. Then, we let each arm compute its own motion plan prefix, which it shares with the other arms, which treat it as a dynamic obstacle. Finally, we add a dynamic priority scheme. The new algorithm, MR-STORM, demonstrates clear empirical advantages over SOTA algorithms when operating with both static and dynamic obstacles.
Paper Structure (23 sections, 15 equations, 2 figures, 5 tables, 1 algorithm)

This paper contains 23 sections, 15 equations, 2 figures, 5 tables, 1 algorithm.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Single-robot motion planning
  4. Multi-robot motion planning.
  5. Background
  6. The formal motion-planning problem
  7. STORM: sampling-based MPC (MPPI)
  8. Decoupled STORM for Multi-Arm MP
  9. Handling Dynamic Obstacles
  10. From Dynamic Obstacles to Arms
  11. The MR-STORM Algorithm
  12. Empirical Evaluation
  13. Setup
  14. Reaching
  15. Following
  16. ...and 8 more sections

Figures (2)

  • Figure 1: Tasks Setups. A. Bin-Loading initial state. Colored cubes: picking positions. B. Reaching-Hard at initial state, level 5. Cuboids are dynamic and static obstacles. C. Following task, level=2. (one obstacle). Goals are mostly around the arms.
  • Figure 2: Average normalized task (circles) and collision (square) scores with one standard deviation. Color code on the bottom. Task score scale on the left. Collision to the left. Dotted lines are the 0 baseline, representing MRS(400,5). For reaching: full circle/square for easy and empty for hard.