David Rother, Thomas Weisswange, Jan Peters, "Summary: Disentangling Interaction using Maximum Entropy Reinforcement Learning in Multi-Agent Systems", AAAI 2023 Fall Symposia: Agent Teaming in Mixed-Motive Situations, 2023.

Research on multi-agent interaction involving both artifi- cial agents and humans is still in its infancy. Current ap- proaches often focus on collaboration-centered human be- havior or a limited set of predefined situations, potentially limiting their efficacy in ”coexistence” environments. These are scenarios likely to arise in future deployments of robots in human-inhabited spaces, where interactions won’t always align with predefined models of collaboration. To address this, we present a novel framework that disentangles inter- action and task-solving into distinct learning challenges, sub- sequently blending the resulting policies during inference. By employing maximum entropy reinforcement learning, we de- velop impact-aware agents and ensure the training cost scales linearly with the number of tasks. Our method proposes a weighting function to align interaction action distributions with the original task distribution. We demonstrate the ef- fectiveness of our framework in addressing scalability, task- solving, and collaboration opportunities in a novel cooking environment. This research paves the way for more intricate robot, multi-human interactions, representing a new learning paradigm.