PhD Dissertation Proposal: Shangqun Yu, A System-Level Approach to Agile and Robust Legged Locomotion in Human Environments
Content
Speaker:
Shangqun Yu
Abstract:
Humanoid robots for disaster response must move dynamically and robustly across rough terrain. This dissertation takes a system-level view, studying how perception, contact decisions, motion control, and hardware design together make this possible. First, on a quadruped guide dog robot, I show what model-based control alone can achieve: real-time nonlinear model predictive control (MPC) and whole-body impulse control (WBIC) produce gentle foot contacts and stable locomotion on slippery surfaces and stairs, with up to 10 dB less walking noise than the robot’s default controller. Second, because optimization struggles to decide when and where to make contact, I train a reinforcement learning policy to choose a humanoid’s gait, footholds, and velocity from terrain observations, while MPC and WBIC generate dynamically feasible whole-body motion. This hierarchy enables walking, jumping, and leaping across discrete terrain, and the same policy transfers to three simulated humanoids without retraining. Finally, the proposed work extends this decomposition to highly dynamic running over rough terrain, where predefined gaits no longer suffice and learned controllers succeed only if simulation captures the hardware faithfully. To this end, I designed the motor controllers and firmware for DASH, an open-source humanoid capable of running at more than 5 m/s. Because the hardware is fully customizable, I can identify its actuator dynamics and communication latency for the training simulation and raise the motor current limits to unlock the actuators’ full torque. I will then train a fully learned hierarchy on a dataset of continuous, highly dynamic motion over rough terrain: a motion generator uses onboard perception to produce whole-body motions, and a general, terrain-aware motion tracker executes them on DASH.
Advisor:
Donghyun Kim