Seminars

Although bipedal locomotion on flat and structured terrains has seen significant progress through methods ranging from reduced-order models to learning-based methods, navigating confined and unstructured environments remains a critical challenge for humanoid robots. This dissertation introduces novel planning strategies to generate controlled, safe, and computationally efficient humanoid locomotion in these complex settings. I will introduce a multi-stage whole-body planning framework that generates dynamically feasible motions while enforcing environment and self-collision avoidance. This is done by strategically introducing convex relaxations to plan the paths of the robot’s torso, hands, knees, and feet as a set of constrained particles. The resulting paths are then refined by including volume-aware and differentiable collision avoidance constraints that increase the reliability of these guiding paths. As a last step, dynamically feasible motions are efficiently found by focusing on motions near the aforementioned collision-free paths. This last step uses a two- pass optimization strategy to achieve dynamically, physically, and morphologically
consistent motions.
As Model Predictive Control often governs the mid-level control layer in locomotion, I also present an Adaptive Horizon MPC strategy that maintains the performance of a long-horizon MPC with approximately half the computational effort by leveraging a Neural Network to approximate the solution to a Bilevel Optimization problem searching for the optimal prediction horizon in real time.
In summary, this dissertation presents a Whole-Body Planner that synthesizes a diverse set of dynamically feasible locomotion strategies that conform to the humanoid morphology to traverse confined environments exceeding those posed by NIST standards.