Cooperative Multi-Agent Mobile Manipulation using Learned Dynamical Systems

Saray Bakker, Wendelin Boehmer, Javier Alonso-Mora

Keywords: Multi-agent Collaboration, Learned Dynamical Systems, Mobile Manipulation

TL;DR: A framework for safe and interpretable multi-agent collaboration in mobile manipulation by building upon independently learned single-agent dynamical systems.

Abstract:

We propose a framework for safe and interpretable multi-agent collaboration in mobile manipulation leveraging independently learned single-agent dynamical systems. Each robot is equipped with a stable task-space vector field defined on $SE(3)$, which policy we preserve, while enabling collaboration through velocity scaling. Unlike prior collaborative learning approaches that rely on predicting key poses and delegating safety-critical reasoning to low-level planners, our method incorporates whole-body collision avoidance directly into the policy while remaining close to the original dynamical systems.
We formulate multi-agent coordination as a finite-horizon optimization problem that adapts execution speed, without modifying motion direction, to satisfy task constraints and ensure collision-free behavior. The proposed framework is parallelized with JAX and evaluated in simulation on two mobile manipulators, demonstrating safe execution and improved coordination across both illustrative and randomized scenarios.

Submission Number: 18