Stochastic Dynamic Games in Belief Space

Information gathering while interacting with other agents under sensing and motion uncertainty is critical in domains such as driving, service robots, racing, or surveillance. The interests of may be at odds others, resulting a stochastic noncooperative dynamic game. Agents must predict others’ future actions without communication, incorporate their into these predictions, account for noise information gathering, consider what reveal. Our solution uses local iterative programming Gaussian belief space to solve game-theoretic continuous POMDP. Solving quadratic game the backward pass belief-space variant linear-quadratic control (iLQG) achieves runtime polynomial number linear planning horizon. algorithm yields feedback policies our robot, predicted agents. We present three applications: Active surveillance, guiding eyes blind agent, autonomous racing. win 44% more races than theory 34% planning.