SUMMARY

Advancement in bipedal robot locomotion has paved the way for their potential mainstream deployment. However, real-world deployment presents significant challenges, such as locomotion and navigation safety in uncertain environments. Deploying bipedal systems in such environments requires a comprehensive approach that accounts for uncertainties through high-level task planning and low-level motion planning. This thesis explores a full-stack approach for bipedal navigation in real-world settings, focusing on enhancing safety and navigation capabilities for effective operation in complex scenarios like simulated warehouses, social settings, and outdoor terrains. This thesis aims to provide formal guarantees or reasoning about safety, robustness, and real-time performance.

The thesis investigates integrated task and motion planning for bipedal robots in a partially observable environment with dynamic obstacles. We design a hierarchical framework that integrates low-level locomotion safety specifications into a formal high-level linear-temporal-logic synthesis to provide safety guarantees on the execution of high-level commands, and task completion. Our interconnected framework achieves simultaneous safe locomotion and navigation. We demonstrate the viability of the framework on reach-avoid, stair climbing, and pick-and-place tasks as a case study for mock warehouse deployment.

The thesis also presents the Social Zonotope Network (SZN), a zonotope-based method to predict future pedestrian reachable sets and plan a socially acceptable reachable set for the robot. A SZN-based model predictive controller optimizes over the output of the neural network, with a novel cost function designed to encourage the generation of socially acceptable trajectories. Our framework produces efficient and socially acceptable motion plans for our bipedal robot system Digit.

Finally, this thesis will propose a framework for traversing an environment with uncertain terrain and uncertainty induced by the discrepancy between reduced-order models and the full-order model of a bipedal robot. Given a rough estimate of an elevation map of the environment obtained via Guassian process (GP) regression, and a GP model describing the robot dynamics modeling mismatch, the framework aims to leverage the locomotion capabilities of bipedal systems in traversing rough terrains while optimizing reduced-order-model-based motion plans for full-order model deployment.