Risky Robotics: Developing a Practical Solution for Stochastic Optimal Control
MetadataShow full item record
Risk is a ubiquitous aspect of control and path planning for robots operating in unstructured real‐world environments. Nevertheless, humans still far surpass robots in their ability to evaluate complex tradeoffs under uncertainty through risk analysis and subsequent decision‐making. Many traditional approaches to the stochastic optimal control problem, such as Partially Observable Markov Decision Processes (POMDP’s), suffer from the curse of dimensionality and become computationally intractable in many real-world scenarios. In this seminar, a new class of stochastic control algorithms is proposed that makes use of emerging high‐performance computing devices, specifically GPUs, to perform real‐time uncertainty quantification (UQ) as part of a feedback control loop. These algorithms propagate the time‐varying probability density of the robot state and optimize control actions with respect to accuracy, obstacle avoidance, and other criteria. Key to practical implementation of these algorithms is the fact that many UQ algorithms can be parallelized; thus they can leverage emerging embedded high‐throughput devices for real‐time or near real‐time execution. Following an overview of the general formulation of these stochastic control algorithms, examples are provided in the form of autonomous parafoil and quadrotor flight controllers that make use of real‐time uncertainty analysis for obstacle avoidance in constrained environments. Recent experimental flight tests using embedded GPUs show that a strong coupling between UQ and optimal control offers a practical solution for risk mitigation by autonomous systems.
- IRIM Seminar Series