Multi-Strategy Deployment-Time Learning and Adaptation for Navigation under Uncertainty

Abstract

We present an approach for performant point-goal navigation in unfamiliar partially-mapped environments. When deployed, our robot runs multiple strategies for deployment-time learning and visual domain adaptation in parallel and quickly selects the best-performing among them. Choosing between policies as they are learned or adapted between navigation trials requires continually updating estimates of their performance as they evolve. Leveraging recent work in model-based learning-informed planning under uncertainty, we determine lower bounds on the would-be performance of newly-updated policies on old trials without needing to re-deploy them. This information constrains and accelerates bandit-like policy selection, affording quick selection of the best-performing strategy shortly after it would start to yield good performance. We validate the effectiveness of our approach in simulated maze-like environments, showing improved navigation cost and cumulative regret versus existing baselines.