Safe multi-agent reinforcement learning for multi-robot control

A challenging problem in robotics is how to control multiple robots cooperatively and safely real-world applications. Yet, developing multi-robot methods from the perspective of safe multi-agent reinforcement learning (MARL) has merely been studied. To fill this gap, study, we investigate MARL for on cooperative tasks, which each individual robot not only meet its own safety constraints while maximising their reward, but also consider those others guarantee team behaviours. Firstly, formulate as a constrained Markov game employ policy optimisation solve it theoretically. The proposed algorithm guarantees monotonic improvement reward satisfaction at every iteration. Secondly, approximations theoretical solution, propose two gradient methods: Multi-Agent Constrained Policy Optimisation (MACPO) MAPPO-Lagrangian. Thirdly, develop first three benchmarks—Safe MuJoCo (Safe MAMuJoCo), Safe Robosuite MARobosuite) Isaac Gym MAIG) expand toolkit research communities. Finally, experimental results benchmarks indicate that our can achieve state-of-the-art performance balance between improving satisfying compared with strong baselines. Demos code are available link (https://sites.google.com/view/aij-safe-marl/).2