Terrain-Aware Robust Control for Quadruped Robots
Loading...
Authors
Lokhande, Sanket
Issue Date
2025
Type
Dissertation
Language
en_US
Keywords
Alternative Title
Abstract
In recent years, quadruped robots have captured imagination of both the public and researchers alike. Their capability of traversing uneven terrains surpass their wheeled counterparts. Utilizing this capability of a quadrupedal robot necessitates locomotion frameworks that balance agility, robustness, and adaptability under uncertainty. This thesis presents a unified framework for context-driven locomotion intelligence in torque-controlled quadrupeds, emphasizing minimal reliance on external sensing while leveraging structure in physical interaction. At the core of the framework is a ground contact force-centric predictive control pipeline that fuses coarse terrain priors with proprioceptive feedback to anticipate contact opportunities, along with the ability to react to disturbances post hoc. To support this, we introduce a scalable model abstraction that bridges rigid-body dynamics with contact force estimation, enabling predictive control of body posture and ground reaction forces without full terrain observability. A second layer of the architecture incorporates a low-dimensional gait manifold, learned through structured exploration, which enables efficient switching between dynamic gait models based on terrain features and task constraints which are learnt offline prior to deployment. Finally, we introduce a cross-domain transfer strategy that fine-tunes control parameters learned in simulation using a physics-aware adaptation policy, reducing the sim-to-real gap without manual retuning. We validate our methods on both simulated and physical quadruped platform Unitree A1. Experimental results show improved robustness in locomotion across discontinuous, and partially observable terrain, while maintaining control efficiency and stability.