From Models to Data: Toward a Unified Framework for Agile and Safe Bipedal Locomotion

Achieving agile, efficient, and robust locomotion in bipedal robots remains a grand challenge of robotics. Traditional model-based control methods are theoretically grounded but are often sensitive to model mismatch and state-estimation uncertainty, limiting their adaptability to real-world environments. Conversely, data-driven approaches such as reinforcement learning produce remarkable behaviors but often lack interpretability, require non-trivial reward shaping, and raise safety concerns. This thesis bridges these two paradigms through a unified framework that begins with model-based behavior synthesis and culminates in data-driven adaptation. The first part focuses on constructing walking behaviors and controllers using reduced-order models of locomotion. A hierarchy of planners and controllers is developed to enable robust walking for flat-footed and multi-domain gaits, as well as safety-critical locomotion over constrained footholds such as stairs and stepping stones. Additionally, this work introduces RoMoCo, a modular open-source architecture, a modular open-source architecture designed to unify reduced-order planning, output synthesis, and whole-body control across multiple bipedal platforms. Building on this foundation, the second part introduces data-driven mechanisms that enable robots to improve and personalize their behaviors through various forms of data. Episodic data collected during repeated executions are used to correct modeling errors and reduce constraint violations. Human preference data facilitates automatic gain tuning through interactive feedback. Online robot data enables adaptation of reduced-order models by learning step-to-step dynamics directly from real executions. Finally, large-scale simulation data support a reinforcement-learning framework designed for hardware deployment, where model-guided rewards enable efficient training and introduce perception inputs, yielding policies capable of dynamic stepping-stone traversal on real robots. Together, these contributions form a progression from theoretically grounded model-based control to data-enabled adaptation, demonstrating that reduced-order models and data-driven learning are complementary. Their integration enables bipedal robots such as Cassie and G1 to walk safely, robustly, and efficiently across diverse terrains, marking a step toward human-level agility in legged locomotion.