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

Jasraj Budigam

Abstract:

Embodied intelligence is trending toward multimodal foundation models and hybrid planners that promise open-world generalization. We study a complementary problem: rapid personalization of navigation to diverse user motor capabilities without heavy models or opaque policies. We present a capability-aware meta-learning method for grid navigation that learns a reusable prior for tabular Q-learning. User capability enters the reward through penalties on turning and backtracking, and motor noise is modeled as probabilistic action slip. At adaptation time the agent initializes from the learned prior and continues standard Q-updates.

We evaluate two regimes: Easy $8 \times 8$ and Hard $12 \times 12$ with slip and capability costs. Adaptation uses 200 episodes in Easy and 60 in Hard. Results aggregate 24 test environments and 10 seeds. We report 95% bootstrap confidence intervals and one-sided paired $t$-tests for $p_{\text{better}}$.

On Hard tasks, success rises from $0.083$ to $0.833 ;(p = 2.4 \times 10^{-7})$. Final return improves from $-499.925$ to $-125.729 ;(p = 3.18 \times 10^{-7})$. Path efficiency increases from $0.083$ to $0.583 ;(p = 1.23 \times 10^{-7})$. Steps on success drop from $131.000$ to $45.450$, and SPL increases from $0.023$ to $0.634$. In Easy tasks the pretrained and scratch agents reach identical final performance.

These results show that capability-aware priors enable rapid personalization while preserving the transparency of tabular methods. The approach runs on a CPU-only laptop with minute-scale training, which supports deployment on resource-constrained assistive platforms and complements foundation-model pipelines in open-world autonomy.