MACT: Multi-agent Collision Avoidance with Continuous Transition Reinforcement Learning via Mixup.

Autonomous collision avoidance is a critical component in various multi-robot applications. While deep reinforcement learning (RL) has demonstrated success in some robotic control tasks, it remains challenging to apply it to real-world multi-agent collision tasks due to poor sample efficiency. The limited amount of transition data available and its strong correlation with multi-agent task characteristics have restricted recent research in this area. In this paper, we propose Multi-agent Collision Avoidance with Continuous Transition Reinforcement Learning via Mixup (MACT) to address these challenges. MACT generates new continuous transitions for training by linearly interpolating consecutive transitions. To ensure the authenticity of constructed transitions, we develop a discriminator that automatically adjusts the mixup parameters. Our proposed approach is evaluated through simulations and real-world swarm robots consisting of E-pucks, demonstrating its practical application. Our learned policies outperform existing collision avoidance methods in terms of safety and efficiency.