End-to-end reinforcement learning for multi-agent continuous control

In end-to-end reinforcement learning, an agent captures the entire mapping from its raw sensor data to actuation commands using a single neural network. End-to-end reinforcement learning is mostly studied in single-agent domains, and its scalability to multi-agent setting is under-explored. Without effective techniques, learning effective policies based on the joint observation of agents can be intractable, particularly when sensor data perceived by each agent is high-dimensional. Extending the multi-agent actor-critic method MADDPG, this paper presents Rec-MADDPG, an end-to-end reinforcement learning method for multi-agent continuous control in a cooperative environment. To ease end-to-end learning in a multi-agent setting, we proposed two embedding mechanisms, joint and independent embedding, to project agents' joint sensor observation to low-dimensional features. For training efficiency, we applied parameter sharing and the A3C-based asynchronous framework to Rec-MADDPG. Considering the challenges that can arise in real-world multi-agent control, we evaluated Rec-MADDPG in robotic navigation tasks based on realistic simulated robots and physics enable environments. Through extensive evaluation, we demonstrated that Rec-MADDPG can significantly outperform MADDPG and was able to learn individual end-to-end policies for continuous control based on raw sensor data. In addition, compared to joint embedding, independent embedding enabled Rec-MADDPG to learn even better optimal policies.