Learning State Switching for Multi-sensor Integration

This paper takes a fresh approach towards multisensor multi-agent integration for achieving improved performance using a control theoretic view of "state-switching". The state-switching problem is formulated as a multi-agent reinforcement learning task for maximizing expected payoffs over time solved using a value iteration algorithm. Specifically, we use a problem of tracking an unknown object with unknown (motion) dynamics using manipulators, defined based on the well-known sawyer one-handed manipulator. We demonstrate that our multi-agent reinforcement learning based state switching algorithm shows superior performance compared to using individual sensors. Our trained agent is also further validated by transferring from simulation to a real experimental setup.