
Robots in Dynamic Tasks: Learning, Risk, and Safety
Joel Burdick of Caltech presents a Stanford seminar on controlling robots that operate in environments too complex to model in advance. He opens with two hardware systems, SQUID, a ballistically launched self-stabilizing drone, and PARSEC, an aerial manipulator that anchors sensor modules to cliff faces, then shows how Koopman spectral techniques can learn fluid-structure interactions like ground effect and feed them into a real-time nonlinear model predictive controller. The second half turns to risk: Burdick explains coherent risk measures and describes how risk-aware terrain analysis helped legged and wheeled robots in the DARPA Subterranean Challenge. He then introduces on-line methods for learning the dynamics of unknown moving obstacles, using distributionally robust chance constraints, and a concept called risk surfaces that let a drone adapt to wind disturbances after just 45 seconds of data collection. Timestamped chapters track the progression from hardware design through control theory to field results.