Robot Learning With Crash Constraints

In the past decade, numerous machine learning algorithms have been shown to successfully learn optimal policies control real robotic systems. However, it is common encounter failing behaviors as loop progresses. Specifically, in robot applications where undesired but not catastrophic, many struggle with leveraging data obtained from failures. This usually caused by (i) failed experiment ending prematurely, or (ii) acquired being scarce corrupted. Both complicate design of proper reward functions penalize this letter, we propose a framework that addresses those issues. We consider violate constraint and address problem crash constraints, no upon violation. The no-data case addressed novel GP model (GPCR) for combines discrete events (failure/success) continuous observations (only success). demonstrate effectiveness our on simulated benchmarks jumping quadruped, threshold unknown priori. Experimental collected, means constrained Bayesian optimization, directly robot. Our results outperform manual tuning GPCR proves useful estimating threshold.