Trajectory Optimisation Inspired Design for Legged Robotics

• Main Author: Fisher, Callen
• Other Authors: Patel, Amir
• Format: Doctoral Thesis
• Language: English
• Published: Faculty of Engineering and the Built Environment 2021
• Subjects: Electrical Engineering
• Online Access: http://hdl.handle.net/11427/33715

Control of legged robots is a non-trivial task, especially when looking at aperiodic (non-steadystate) manoeuvres such as rapid acceleration and deceleration. Observing nature, animals are seen to effortlessly perform these rapid transient manoeuvres, however, robust walking is still considered a complex task in the robotics literature. For robots to successfully explore the unknown world outside of the laboratory environment, these transient manoeuvres need to be thoroughly understood and mastered. As controlling and designing legged robots is an extremely complex task, this thesis argues that trajectory optimisation methods can be employed to make various aspects of design more tractable. First trajectory optimisation methods were employed to determine how legged robots should accelerate. In nature animals are seen to leap straight into the desired gait, however, in the robotics literature, robots often perform multiple discrete gait transitions, or accelerate extremely slowly. The results of the study revealed that the optimal method to accelerate was to launch straight into the desired gait (for both bipeds and quadrupeds), with a sliding mass template model emerging for all results. Another discrepancy between the literature and nature is that animals have active spines which have been shown to aid in rapid locomotion tasks. In the robotics literature a number of spines exists, however, which is the optimal morphology for transient manoeuvres? Using Monte Carlobased trajectory optimisation methods, the rigid, revolute and prismatic spine morphologies were compared, with results showing the prismatic spine was the optimal configuration for long-time-horizon tasks. Due to transient locomotion requiring accurate and complex whole-body models, resulting in computationally expensive optimisation problems, an “optimisation-inspired” approach (akin to “bioinspiration”) was taken to identify heuristics and trends for a monopod robot. Initially, optimal trajectories for a two link leg monopod were analysed. Interestingly, during the stance phase, the axial force of the leg behaved in a “bang-bang” like fashion with fine hip torque control. Furthermore the aerial phase showed correspondence to the popular Raibert style controller. This resulted in the development of a hybrid pneumatic-electric monopod robot as a test-bed. Trajectories were then optimised for the robot to determine if these heuristics held. From these results, a stance phase PD controller for the hip actuator was developed and simulated under disturbances to test robustness. The resulting controller and heuristic was successfully tested on the platform, performing a long-time-horizon motion, which included transient phases of acceleration and deceleration.