STRIDE: Sophisticated Temporal Robots for Inference-guided Decision Exploration

Limb utilization in legged robots enables these platforms to navigate in more complex terrains compared to wheeled or tracked platforms. However, state-of-the-art legged robot locomotion is still far from matching the performance of animals, which can run or walk at brisk paces over very complex terrains. Aptima proposes to build and evaluate STRIDE: Sophisticated Temporal Robots for Inference-guided Decision ExplorationŚan adaptive, learning locomotion controller that will enable quadruped robots to quickly accomplish goals without slipping, falling, or getting stuck in environments. STRIDE solves the control problem via probabilistic inference and minimization of goals and model uncertainties, resulting in Bayesian-optimal exploration of leg action policies. Our solution offers several unique benefits for Army infantry units and special operations forces: (1) continually updates its understanding of the world, enabling quicker execution of tasks by identifying efficient actions with fewer interactions; (2) utilizes multimodal sensing to ensure robustness to occlusions and ground cover; and (3) facilitates faster learning of new skills through exploratory behaviors. By incorporating real-time learning and adaptation, STRIDE eliminates the need for offline training, enhancing the robot's capabilities in dynamic and unfamiliar environments, thus empowering quadruped robots to serve as reliable and effective teammates for Warfighters.