Joint mapping and planning in multi robot systems

Abstract

The field of Multi Agent Reinforcement Learning The field of Multi Agent RL can model and address compplex interactions among multi-decision makers in a shared environment. There is however, a need of efficient coordination and communication among agents. Centralized Training methods are ineffcient for large scale, dynamic deployments. There is work to be done on coordination amongst decentralized agents. However, most works in MARL assume ideal network conditions, where real life wireless network constraints are not considered. Investigating decentralized learning ofagent policy (for action and communication) in realistic network conditions and developing robuststrategies enables us to deploy these algorithms in real life use cases.We explore using Decentralized Multi Agent Deep RL techniques to the problem of Multi Agent systems being able to jointly navigate and collectively map environments. Advances in this shalllead to developement of swarms that can monitor infrastructure, environment ecosystems, post-disaster assessments, indoor military operations, etc.While literature in Distributed SLAM looks into agents using handcrafted features with predefined communication protocols, they do not consider navigation being adaptive to the mapping process. We are further looking into compression of map data that is exchanged between agents,the (learned) communication protocols and integrating it with Active SLAM.