Attention-Based Spatial Encoding for Multi Agent Coordination

Coordinating groups of aircraft and unmanned aerial vehicles to reach target locations without collisions is crucial for deploying decentralized groups of vehicles in the real world. A key challenge is coordinating a variable number of agents without the need for agents to share information about their intended routes, reducing communication requirements to a minimum. Recent advances in natural language processing demonstrate the efficacy of attention mechanisms in Transformer neural networks to handle long sequences of varying sizes at unmatched performance levels. Based on this insight, we propose attention-based spatial encoding combined with deep reinforcement learning to solve the outlined coordination problem. We demonstrate state-of-the-art interpretability, scalability, and performance in simulations compared to a competitive baseline that uses Long-Short-Term Memory modules for spatial encoding.