Learning to Walk - Optimal Gait Synthesis and Online Learning for Terrain-Aware Legged Locomotion

Abstract: 

The goal of the proposed research is to advance the science of cyber-physical systems by more explicitly tying sensing, perception, and computing to the optimization and control of physical systems whose properties are variable and uncertain. The CPS platform to be studied is that of a bipedal robot locomoting over granular ground material with uncertain physical properties (sand, gravel, dirt, etc.). Legged robots form a class of cyber-physical systems characterized by a tight coupling between the cyber component (the robot) and the physical world (the ground). Only through the controlled interaction of the robot’s feet with the physical environment can a legged robot achieve locomotion. When translated to the CPS platform, the research goal is to improve the perception and control of legged locomotion over granular media for the express purpose of achieving robust, adaptive, terrain-aware locomotion for legged robots. The current state-of-the-art for bipedal locomotion focuses primarily on solid ground (whether flat, sloped, stepped, or irregular), yet the natural and engineered worlds consist of a larger variety of terrain types. Applying existing bipedal walking strategies to these terrain types fails due to the (unmodeled) dynamics associated with the foot-substrate interaction. 

The proposed work builds on existing expertise in granular media mechanics, optimal control for periodic systems, bipedal locomotion, neuro-adaptive control, and real-time, real-world computation. We hypothesize that simple models with decent predictive performance and low computational overhead are sufficient for the optimal control formulations as the compute-constrained adaptive subsystem will both learn and classify the peculiarities of the terrain online. The granular interaction mechanics are central to the cyber-physical formulations, with the physical experiments guided by the capabilities of the cyber component, and informing the design of the cyber components. The main research objectives will involve: [1] a validated co-simulation platform for legged robot movement over granular media; [2] terrain-dependent, stable gait generation and gait transition strategies via optimal control; [3] online, compute-constrained learning of granular interactions for adaptation and terrain classification; and [4] validated contributions using experimental testbeds involving variable and unknown (to the robot) granular media.