Robust policy search with applications to safe vehicle navigation

This work studies the design of reliable control laws of robotic systems operating in uncertain environments. We introduce a new approach to stochastic policy optimization based on probably approximately correct (PAC) bounds on the expected performance of control policies. An algorithm is constructed which directly minimizes an upper confidence bound on the expected cost of trajectories instead of employing a standard approach based on the expected cost itself. This algorithm thus has built-in robustness to uncertainty, since the bound can be regarded as a certificate for guaranteed future performance. The approach is evaluated on two challenging robot control scenarios in simulation: a car with side slip and a quadrotor navigating through obstacle-ridden environments. We show that the bound accurately predicts future performance and results in improved robustness measured by lower average cost and lower probability of collision. The performance of the technique is studied empirically and compared to several existing policy search algorithms.