Reinforcement Learning for Assessing Route Instruction Usability in Complex Indoor Spaces

Wayfinding in complex indoor spaces is challenging, particularly when route instructions are incomplete or ambiguous. While humans often successfully navigate under such conditions by leveraging past experiences, computationally modeling this adaptive skill remains an open research area. This paper addresses the problem of assessing the usability of incomplete route instructions in unknown indoor environments. The objective is to determine if Reinforcement Learning (RL) agents, trained in diverse settings, can acquire transferable wayfinding skills to navigate effectively despite missing information. We demonstrate a novel computational approach using RL, specifically Proximal Policy Optimization (PPO), to model the acquisition and transfer of wayfinding skills. Agents are trained via a curriculum learning approach in simulated text-based indoor environments of varying complexity and with different turn-based instruction grammars. Their ability to navigate with both complete and incomplete instructions is then evaluated in seen and unseen environments. Our results show that RL agents successfully learn to navigate even with incomplete instructions, significantly outperforming random agents. Agents trained on diverse environments generalize well to novel settings, although performance decreases with higher environmental complexity and finer-grained instruction grammars when instructions are incomplete. Step-cost reward structures yielded better learning outcomes. This research offers a probabilistic framework for quantifying route instruction usability. The framework reframes route evaluation from deterministic error checking to probabilistic risk assessment. It lets designers gauge when `imperfect’ instructions remain usable, optimise wording and grammar, and develop adaptive indoor navigation, evacuation, or accessibility aids. By quantifying skill transfer across layouts, it also provides an experimental testbed for cognitive theories of heuristic learning and spatial knowledge acquisition