GCMRBench: Goal-Conditioned Multi-Robot Environments and Benchmarks for Advancing Offline Multi-Agent Reinforcement Learning

Chenxing Li, Chin-Jui Chang, Yinlong Liu, Zijian Ma, Jan R. Seyler, Shahram Eivazi

Keywords: Offline Multi-Agent Reinforcement Learning, Benchmarks, Dual Robot, Goal-Conditioned Environments

TL;DR: We present a novel simulated platform for multi-agent reinforcement learning and comprehensive benchmarks.

Abstract: Research in multi-agent reinforcement learning (MARL) has been focused on developing algorithms to solve issues arising from agents' diverse goals, collaboration and competition needs in complex environments. The extension of these algorithms to Offline MARL (OMARL), along with the utilization of large-scale offline data, is increasingly proposed as a potential step toward achieving safe, efficient, and rapid deployment in real-world scenarios. However, the existing body of research often trains and evaluates OMARL in environments predominantly designed for game scenarios. As a result, the potential of utilizing OMARL remains an open question in domains such as robotics. To bridge this gap, here we introduce GCMRBench, a goal-conditioned multi-agent simulation environment tailored for dual-arm robotic tasks. Our benchmark includes five categories of multi-agent tasks: 1) Cooperation, 2) Competition, 3) Multi-Goal, 4) Transition, and 5) Hybrid systems. It includes a comprehensive suite of 22 tasks, 56 datasets, 6 representative OMARL algorithms, and an example of a real-world deployment system. By open-sourcing our environment, datasets, and algorithms, we provide a platform and baseline for the development and evaluation of offline multi-agent algorithms, thereby accelerating the deployment of these algorithms in practical robotic applications.

Area: Robotics and Control (ROBOT)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1192