Bimanual Multi-Agent Robot Learning Benchmark

I built a bimanual multi-agent manipulation task to study how reinforcement learning methods handle coordination in a shared robotic scene. I also evaluated two existing ManiSkill tasks so the new task could be compared against established benchmarks.

The work connects task design in ManiSkill with training and evaluation in HARL. I trained RL baselines from scratch, including PPO, MAPPO, HAPPO, and SAC, then compared how single-agent and multi-agent approaches behave across the task suite.

A longer writeup will cover the task design choices, reward shaping, evaluation protocol, and what the baselines revealed.