Decentralized Multi-Robot RL: From Isaac Lab/Isaac Sim to Physical Robots

[Abddo9]

I'd like to share our recent research on decentralized multi-agent reinforcement learning for multi-machine tending.

Our work introduces FMAPPO (Feature-fusion Multi-Agent PPO), combining LiDAR observations with task-specific state information to learn joint task assignment and navigation policies.

We use Isaac Sim/Lab for high-fidelity simulation and evaluate the learned policy on physical mobile robots through a ROS 2-based deployment pipeline.

h_out.mp4

📄 Paper: https://arxiv.org/abs/2609.14567
More Videos of the real deployment: https://anonymouspapers123.github.io/FMAPPO/

The implementation is not publicly available yet, but I plan to share the code after the paper is accepted. In the meantime, I'd welcome discussion with the community about sim-to-real RL, multi-robot coordination, and realistic robot-learning environments.

[vick-yu]

Looking forward to the code release when the paper clears review. This kind of multi-robot coordination in industrial settings is exactly the use case Isaac Sim was built for -- the high-fidelity physics and ROS 2 bridge make it a natural fit for sim-to-real pipelines like yours.