Abstract
Quadruped robots have strong terrain adaptability and high mobility, making them highly promising for applications in industrial inspections, emergency rescue, and other fields. As a result, they have become a hot topic in robotics research. Locomotion control is a fundamental function of quadruped robots; however, the highly nonlinear dynamics of quadruped robots and the complexity and diversity of external operating environments make locomotion control extremely challenging. This dissertation aims to systematically address the difficulties of quadruped robot locomotion control by utilizing and developing theories of model predictive control, numerical optimization, robust optimization, and reinforcement learning. The specific research contents are as follows:To address the real-time issue of quadruped robot locomotion control, a fast model predictive control algorithm based on the online active set method is proposed. First, a quadratic program model for quadruped robot motion control is constructed. Then, two numerical optimization techniques, global inactive constraints and dimension-invariant warm start, are proposed for the real-time online solving this quadratic program. An online active set strategy based on an approximate quadratic program is designed. Finally, simulations and real-world experiments verify that the algorithm effectively improves the solution efficiency and enhances the real-time performance of locomotion control.
To handle the robustness issue of quadruped robot locomotion control, a min-max model predictive control algorithm based on robust optimization is proposed. First, a robust min-max model predictive controller is proposed to address the uncertainties in the body dynamics and external environment in locomotion control. The Lagrangian duality is used to prove that the original problem is equivalent to a convex quadratically constrained quadratic programming problem. Then, a numerical algorithm based on the trisection method is proposed to achieve real-time solutions for this problem. Finally, simulations and real-world experiments verify that the algorithm can effectively deal with challenges such as inaccurate dynamic models and low-adhesion surfaces, thereby improving the robustness of locomotion control.
To deal with the safety issue of quadruped robot motion control, a collision-avoidance model predictive control algorithm based on convex optimization is proposed. A convex optimization-based passable region calculation method is proposed, establishing convex polygon constraints centered on the robot, and designing a model predictive controller based on these convex collision-avoidance constraints. Finally, simulations and real-world experiments verify that the algorithm can ensure that the robot does not collide with obstacles even when given an infeasible reference trajectory, thereby improving the safety of locomotion control.
To achieve autonomous gait decision-making for quadruped robot locomotion, a hierarchical control algorithm based on reinforcement learning and model predictive control is proposed. A Markov decision model for gait switching is established, and the proximal policy optimization algorithm is used to solve the optimal gait switching strategy. Simulations and real-world experiments verify that the algorithm can autonomously decide the optimal gait under different operating conditions, improving the energy efficiency of locomotion control.
| Date of Award | 24 Jun 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Chin Pang HO (Supervisor) & Lijun Zhu (External Supervisor) |