Risk based control in uncertain environments

Abstract

Controllers for dynamic robots which regulate around an optimized trajectory often struggle with reliability in uncertain environments, and making control decisions in real-time. Using a map of risk to the robot's state, a risk gradient controller can be created to find paths back to safety from anywhere in the robot's reachable space. Despite generating a variety of complex and dynamic behaviors, this method suffers from sensitivity to sensor and process noise, as the value of risk associated with the robot's state is not well behaved. In order to correct for this sensitivity, a Stochastic Risk Gradient Controller (RGC) and sample-based Extended Kalman Filter (EKF) are proposed. The Sample Based EKF uses pre-simulated dynamics to generate optimal state and uncertainty estimate up to 10 times faster than an online-simulation. The Stochastic RGC then uses those estimates to calculate more robust control actions in real-time. This framework is applied to a model of a pogo-stick robot and simulated with various levels of sensor and process noise. In trials with relatively large measurement noise, the Stochastic Risk Gradient controller succeeded up to 15% more often than the naive risk gradient controller, while failing up to 8% more often in cases when measurement noise was relatively low.