On Learning Techniques for Optimal Control of Legged Robots
Daniel Felipe Ordoñez Apraez (Ph.D. Student)
In this project, we will focus on the integration of data-driven methods to model-based optimal-control pipelines that formulate the optimization process as a trajectory optimization. The objective is to find application areas for data-driven methods to improve the control pipeline of a legged robot. The tentative areas of focus will be dynamics modeling, system identification, gait features selection, foothold localization, full-body trajectory optimization, and short and long-term planning. The desire to exploit data-driven methods for legged locomotion arises from the need to address several of the pitfalls that traditional receding horizon controllers suffer when operating in unstructured terrains and diverse scenarios, limiting their versatility and widespread implementation. Among these pitfalls we find: (1) Efficiency vs Complexity trade-off: Receding horizon controllers for legged locomotion are required to trade computation time and resources with the quality of the modeling and solution to the optimization problem. This trade-off often leads to sub-optimal control policies, or a very high control bandwidth when compared to animals in nature. (2) Dependency on heuristics and model assumptions: Most state-of-the-art receding horizon controllers depend on modeling assumptions (e.g., linearized or floating-based dynamics, ideal contact models, point contacts) or human heuristics (e.g., CoM angular momentum minimization, fixed gait patterns, gait period) to improve efficiency or ensure convergence of the optimization process. (3) Non-Adaptability: All receding horizon formulations that do not exploit data from experience are fixed control pipelines, having no capacity to adapt to changes in the dynamics of the robot (e.g., kinematic or dynamics changes/disturbances) nor to changes in the interaction with the environment (e.g., contact dynamics, external disturbances).
| Primary Host: | Massimiliano Pontil (Istituto Italiano di Tecnologia & University College London) |
| Exchange Host: | Carlos Mastalli (Heriot-Watt University & Institute for Human and Machine Cognition) |
| PhD Duration: | 01 November 2022 - 01 November 2025 |
| Exchange Duration: | 01 November 2023 - 01 November 2024 - Ongoing |
