Multi-Agent Actor-Critic Model Predictive Control

Multi-Agent Reinforcement Learning Model Predictive Control Drone Swarms Hardware Pending Submission

This work develops multi-agent actor-critic model predictive control (MA-AC-MPC), a framework that combines multi-agent reinforcement learning with model-based control to produce safe, dynamically feasible actions for cooperative robotic teams. The learning policy handles long-horizon, non-differentiable team rewards, while MPC provides fast replanning over short horizons with the physical constraints of each robot embedded directly in the action layer.

The approach is evaluated in multi-agent pursuit-evasion and in a heterogeneous drone-rover landing task. In hardware, MA-AC-MPC achieved repeatable drone landings on a moving omni-wheeled rover with a 100% success rate, compared to 60% for a multi-layer perceptron policy, while also demonstrating robustness to agent mass variations and real-world uncertainty.

Crazyflie drone mounted on an omni-wheeled rover platform MA-AC-MPC architecture diagram Multi-agent pursuit-evasion trajectories with pursuers and evaders Drone and rover cooperative landing trajectory Curriculum learning win rate comparison Mass sweep comparison between MA-AC-MPC and MA-AC-MLP MA-AC-MPC trajectory comparison MA-AC-MLP trajectory comparison