Hi everyone,
I’d like to share our recent research on decentralized multi-agent reinforcement learning for multi-robot task assignment and navigation, developed using NVIDIA Isaac Sim/Isaac Lab and evaluated on physical mobile robots.
📄 Paper: Learning Multi-Agent Task Assignment and Navigation in the Factory: From Simulation to Real Robots
In this work, we investigate how multiple mobile robots can coordinate to service multiple machines, navigate to collect completed parts, and deliver them to designated storage areas without relying on a centralized execution policy.
We introduce FMAPPO (Feature-fusion Multi-Agent PPO), a decentralized reinforcement-learning approach that combines 2D LiDAR observations with task-specific state information to learn task assignment and navigation jointly.
Our workflow builds a complete pipeline from simulation-based policy training and evaluation in Isaac Sim/Isaac Lab to real-robot deployment through ROS 2. We also investigate how the policy responds to different command update frequencies during physical deployment, highlighting some of the practical challenges that arise when transferring learned policies from simulation to real robotic systems.
The goal is to move beyond simulation-only evaluation and better understand the challenges of deploying decentralized robot-learning policies under real-world sensing and control constraints.
The code is not publicly available yet, but I plan to release it after the paper is accepted.
I’d be interested in exchanging ideas with others working on Isaac Lab, reinforcement learning for robotics, multi-robot coordination, and sim-to-real transfer.
More Videos of the real deployment are available here
Thanks for taking a look!