Decentralized learning helps robots avoid crashes without central control
Decentralized Safe Multi-Agent Reinforcement Learning via Predictive Shielding
Artificial IntelligenceMultiagent SystemsRobotics
Summary
When multiple robots work together without knowing much about each other, keeping them safe and efficient is hard. The paper shows how teaching robots to predict and avoid dangers ahead of time helps them adapt better to new situations. The authors designed a method where each robot makes its own decisions while still preventing conflicts, without needing a central boss to control everything. They also came up with a way for robots to solve tie situations quietly without talking. This could make robot teams safer and more flexible in busy environments.
multi-agent systemsreinforcement learningsafety shieldsdecentralized controlpredictive shieldingmodel-based learningfinite horizon Q-learningconflict resolutionlivelocksrobot coordination
Authors
Yacine El Yamani, Hanna Krasowski, Elena Vanneaux
Abstract
Environments are increasingly populated by multiple robots performing independent tasks with limited prior knowledge of each other. Deploying such multi-agent systems presents significant challenges. Specifically, shifts in deployment states compared to training data can lead to poor policy performance and compromised safety. While safety shields exist to mitigate these risks, they are typically reactive, which degrades performance near unseen obstacles,and centralized, limiting their scalability. To address this, we propose a decentralized framework that integrates predictive shielding with model-based finite horizon Q-learning. This approach allows agents to safely adapt their pre-trained policies during deployment. Furthermore, to mitigate livelocks in symmetric scenarios, we introduce a communication- free protocol for conflict resolution