AI-Driven Assistive Robotics: Reinforcement Learning For Disability Support

Abstract

This thesis investigates whether transfer learning can accelerate reinforcement learning for a simulated autonomous wheelchair navigation agent across three progressively complex Unity environments. A staged curriculum was designed, moving from basic empty-room navigation through static obstacle avoidance to dynamic NPC staff and patient comfort constraints. The central finding is that a model initialized from Stage 2 weights outperformed a from-scratch model by 14.4 percentage points in Stage 3, with the entire advantage coming from reduced timeouts rather than better obstacle avoidance. The thesis also documents the iterative reward engineering failures behind each working solution, including a network architecture mismatch that forced a locking policy shaping all subsequent design decisions.