Terrain-Aware Bipedal Locomotion

Abstract

Bipedal robots provide a path forward for autonomous systems to operate seamlessly in human-designed environments. They can traverse terrains compatible with wheeled robots and provide improved access to shelves and cabinets out of reach for quadrupeds and other mobile robots. Though their design offers many advantages over other platforms, the consequential challenge of stability remains a glaring hindrance in realizing their true potential. To this end, our work investigates the evolution of algorithmic strategies that enable these robots to traverse various terrains and actively engage with their surroundings dynamically.

To mitigate the unreliability of walking in steep or slippery environments, we first design and test a terrain-aware foot placement locomotion controller (ALIP-MPC) on a 20 Degree-of-Freedom (DoF) Cassie robot. ALIP-MPC displays improved results compared to foot placement methods that disregard terrain information. The controller is validated in simulation and hardware and performs better than other state-of-the-art foot placement methods.

Subsequently, we extend the ALIP-MPC method for a 30-DoF Digit where perception data is used to provide real-time terrain information online. The proposed method is a multi-stage receding horizon algorithm that utilizes properties of the Angular Momentum Linear Inverted Pendulum (ALIP) model for fast execution speeds. Initial results of the fully integrated locomotion controller are shown in simulation and hardware.

Lastly, we broaden the concept of terrain-aware control to encompass interaction with the environment in diverse whole-body tasks. We develop Kinodynamic Fabrics for reactive whole-body control on a 30-DoF Digit robot. This method integrates optimization fabrics within a whole-body nullspace control schema to achieve a range of motions, including balancing and walking.