Perception and control of robot legged locomotion over variable terrain

Abstract

This thesis investigates the role of foot-ground dynamics in robot legged locomotion. Empirical observation with robot hardware is critical for such investigation due to the complexity of impact dynamics and the infinite variability of real terrain. To this end, we introduce the MIT Super Mini Cheetah, a small, inexpensive quadrupedal robot that can run and walk over a variety of ground types (gravel, tile, carpet, grass, etc.). Experiments with the Super Mini Cheetah demonstrate new terrain-adaptation capabilities for locomotion control over surfaces with varying mechanical impedance. The terrain adaptation methods are shown empirically to increase gait stability and to reduce impact loading on the legs. To the best of our knowledge, the work demonstrates the first example of a running controller that measures changes in stiffness of the ground and modifies its controller within a single stride. This enables the Super Mini Cheetah to run smoothly between hard and soft surfaces. Algorithms to quickly measure ground properties are presented. These measurements are made by observing interaction between robot legs and the ground. We show that the impact event between the foot and the ground is particularly information-rich which enables the Super Mini Cheetah to estimate ground stiffness within 50 milliseconds of beginning a step. These results contribute new knowledge about the hardware and control requirements to provide legged robots with a useful sense of touch, which will greatly enhance the types of terrain that legged robots can safely and gracefully roam.