In multi-robot manipulation, the value of any single robot's action depends on how its teammates will respond. This interdependence makes coordination difficult when robots must act autonomously without a central planner. A new framework, called Cooperative Action–Response Belief Estimation (CARE), addresses this by learning what each robot can expect from others based only on its local history.
CARE builds a model of team-response beliefs—predictions about teammates' future actions—from the information a robot has already observed. This allows a robot to reason about the likely consequences of its own choices before acting, even when it cannot see the full state of the team or the environment.
The approach is designed for decentralized multi-robot manipulation, where communication may be limited and each robot must make decisions independently. By grounding action selection in learned beliefs about teammate behavior, CARE offers a way to improve coordination without relying on centralized oversight or exhaustive state sharing.