Reproduce “Statistical Safety and Robustness Guarantees for Feedback Motion Planning of Unknown Underactuated Stochastic Systems” on OpenResearch

About this paper

Craig Knuth, Glen Chou, Jamie Reese, Joe Moore

Researchers from Johns Hopkins University Applied Physics Laboratory and the University of Michigan developed SAFED, a feedback motion planning method providing statistical safety guarantees for unknown, underactuated, stochastic robotic systems. The approach integrates learned dynamics, Extreme Value Theory for uncertainty quantification, and contraction theory, demonstrating robust performance on quadrotors and ground robots where prior methods failed due to unbounded model errors.

Key insights

• EVT can be effectively adapted to statistically bound model errors in stochastic systems by focusing on the maximum residual over a trusted domain, overcoming the issue of unbounded Lipschitz constants.
• Integrating learned dynamics, statistically verified uncertainty bounds, and contraction theory enables robust feedback motion planning with quantifiable safety guarantees for underactuated systems.
• The proposed framework allows planners to implicitly learn and avoid regions of high uncertainty or risk, adapting safety considerations to varying environmental conditions.

Summary by alphaXiv.