Aerospace Controls Laboratory
Leveraging recent advancements in the safety, availability, and predictability of mobile robots, this project aims to design algorithms that enhance these qualities in collaborative warehouse and delivery robots. One key goal is to advance our state-of-the-art navigation algorithms, which rely on reinforcement learning (RL) and model predictive control (MPC), to handle uncertainties such as imperfect perception and to predict the intent of other agents, allowing for more seamless interactions with humans in pedestrian-dense environments. The complexity of this goal requires integrating learned components into the autonomy stack, driving our second goal: to provide formal safety guarantees for AI-based navigation algorithms in dynamic, multi-agent environments. This involves extending our recent neural network verification algorithms to efficiently compute reachable sets and verify system safety.
This video showcases the agility of the custom-designed STAR (Socially Trained Agile Robot) robot as it navigates through a crowded hallway.