An Incremental Approach for Multi-Agent Deep Reinforcement Learning for Multicriteria Missions

Autonomous mobile robots are increasingly used in scenarios where humans cannot be sent due to risk or difficulty, such as exploration or reconnaissance missions in extreme conditions. These missions require multiple agents to communicate and cooperate through a decentralized network. Thus, a successful operation requires a network topology capable of maintaining connectivity between agents and being resilient to failures. The network must also be able to identify and avoid possible threats. Multicriteria missions, in which several activities must be performed simultaneously, are suitable for Machine Learning techniques such as Multi-Agent Deep Reinforcement Learning (MADRL). We applied Proximal Policy optimization (PPO) and Centralized Training with Decentralized Execution (CTDE) to train a network of 20 agents in learning 4 distinct tasks incrementally: movement and obstacle avoidance, connectivity maintenance, network resilience maintenance, and area coverage enhancement. The resulting model was able to generate such robust network topology and showed promising results for 3 out of 4 tasks.