| dc.description.abstract | Legged robots have long been envisioned as a means of expanding robotic capabilities beyond structured environments, yet achieving high-agility locomotion remains a fundamental challenge. This thesis presents a model-based framework for parkour-style locomotion, enabling robots to execute highly dynamic maneuvers such as jumps, rolls, and flips with precision and robustness. A key challenge in planning these motions is selecting an appropriate dynamic model that balances computational efficiency with physical accuracy. To address this, a model assessment strategy is introduced to determine the simplest model capable of capturing task-relevant dynamics. Even with well-chosen models, solving long-horizon trajectory optimization problems for dynamic motions is computationally demanding. This thesis introduces graduated optimization techniques, which improve solver efficiency and reliability by generating high-quality initial guesses through progressively refined problem formulations. Additionally, a novel formulation of rigid-body dynamics algorithms for systems with kinematic loops accelerates trajectory optimization and simulation. Finally, two control strategies are proposed to execute planned motions on hardware: a model-based tracking controller for real-time adjustments and an imitation learning policy trained on optimal trajectories to enhance robustness. Extensive experiments on hardware validate the framework, demonstrating the successful execution of complex, high-impact locomotion behaviors. By integrating advanced planning, optimization, and control techniques, this work establishes a foundation for high-agility legged locomotion, pushing beyond conventional automation toward real-world, dynamic robotic movement. | |
