Title: Safe Robot Motion Generation With Gaussian Processes
Abstract: Evaluating and deploying machine learning models on physical robot hardware is challenging for a number of reasons. Active data-collection is costly because real-world robots operate at real-time and cannot produce s ufficient training data in a reasonable amount of time. Machine learning models are often randomly initialized which results in unsafe and unpredictable beh aviour in the early phases of training. Interacting with the physical world requires satisfying strict safety and feasibility constraints, w hich are often difficult to quantify or simulate accurately. Probabilistic models, such as Gaussian processes, are excellent candidates for learning tasks in rob otics. Their robustness to measurement noise, ability to model highly non-linear functions, and exceptional data-efficiency makes them ideal for robotics applications. This work explores safety across diverse GP applications, including motion planning, model-based rei nforcement learning, and Bayesian optimization. We present modifications that make these algorithms more practical for execution on real-world robots.
|March 11, 2024 (Mon) 17:00 - 18:00