Partially Supervised Reinforcement Learning for GPS-Denied Navigation

Navigating dynamic environments is a key challenge for autonomous agents, yet most existing research focuses on 3D settings or settings where the agent has full access to relevant semantics. In this work, we propose a learning framework for aerial navigation in the presence of changing dynamics and limited positional information. Specifically, we consider a drone navigation task where a drone at one time has access to GPS location information, which it has now lost and needs to navigate in the same area but at a future time with no GPS signal and differing transition dynamics. To address this, we introduce a novel simulator and algorithm for navigation that provides a similarity map of the agent's position compared to a previously taken reference map of the area. This, combined with other state information, is fed to a feature extractor to calculate a map of the agent's estimated position. The position estimate is provided to a policy network, allowing the agent to navigate. We show how the locational information can be useful for understanding the agent's current perception of its environment and show interesting emergent behaviors of the agent. Our results underscore the importance of robust, sample-efficient learning methods for real-world scenarios where the environment evolves over time and precise localization cannot be assumed.