Meta-Learning for Rapid Adaptation in Assistive Navigation: A Capability-Aware Approach

Jasraj Budigam

Abstract: Assistive navigation systems face a fundamental challenge: they must quickly adapt to the diverse motor capabilities of individual users while navigating different environmental layouts. This work presents a meta-learning approach that learns capability-aware priors using tabular Q-learning across a range of grid-based navigation tasks. We evaluate our approach in two distinct regimes: Easy tasks featuring $8 \times 8$ grids with moderate clutter and no capability costs, and Hard tasks on $12 \times 12$ grids with dense clutter, probabilistic action slip, and capability-shaped reward functions. In both regimes, our agent first trains across multiple source environments to build a robust prior, then rapidly adapts to novel user–environment combinations over just a handful of episodes. Our results demonstrate substantial improvements in the challenging Hard regime, where the pretrained agent successfully reaches the goal in $83.3%$ of test maps compared to only $8.3%$ for agents learning from scratch. We observe dramatic improvements in mean final return (from approximately $-499.9$ to $-125.7$) and path efficiency (from $0.083$ to $0.583$). Importantly, in the simpler Easy regime, both pretrained and scratch agents achieve comparable final performance, demonstrating that our learned prior doesn't hinder adaptation when tasks are straightforward. The method operates with minimal computational requirements, relying on efficient tabular updates and brief adaptation windows, while consistently delivering substantial benefits in challenging, sparse-reward environments, precisely where rapid personalization matters most for assistive navigation.