Embodied object hunt

Abstract

In dynamic and unstructured environments, the ability of autonomous agents to perform Object Goal Navigation (ObjectNav) is critical for a wide range of real-world applications, including assistive robotics, disaster response, and autonomous warehouse management. This thesis builds upon the SPOC (Shortest Path Oracle Clone) framework, which leverages imitation learning on shortest-path trajectories to train agents capable of navigation, exploration, and manipulation using only RGB inputs. Despite its strong performance, SPOC remains vulnerable to failure loops, where agents become stuck in repetitive and unsuccessful actions.

To address this limitation, we propose a novel enhancement that integrates multiple learning strategies, which includes action feedback signals (last_action_success), Low-Rank Adaptation (LoRA), modular backtracking, and failure caching, to improve decision-making, adaptability, and robustness.

Through systematic experimentation, we demonstrated that our approach improved ObjectNav performance, achieving a test success rate of 84.4%, outperforming our tested baseline of 79.6%. Furthermore, our method improved exploration efficiency, reducing the episode length by 20.7% for successful episodes and 59.4% for failed ones.