Failure Prediction from Limited Hardware Demonstrations
Bayesian Decision making and Uncertainity workshop (BDU), NeurIPS 2024, Allerton, 2024
Failure discovery of robotic systems faces a unique challenge with the emergence of sim-to-real failures, that cannot be captured by simulation models, or analytically modeled to replicate in real-world platforms with ease. At the same time, simulation models share vital information about failures that will be replicated in real-world platforms. In this work, we explore a combination of sampling-based techiques and Bayesian Experimental Design (BED) to learn failure prediction models, by utilizing sample-extensive exploration in sim and sample-efficient failure discovery in real systems, supported by failures discovered in simulation. The project website can be found here.
Recommended citation: A. Parashar, K. Garg, C. Fan (2024). "Failure Prediction from Limited Hardware Demonstrations." IEEE proceedings, Allerton 2025.
Download Paper