Research

RABBIT A rigid, five-link, four-actuator planar biped for studying underactuation and impacts in walking.	MABEL A compliant bipedal robot that connected HZD ideas to efficient walking and fast running.
MARLO A 3D robot that moved the agenda from planar demonstrations toward outdoor walking.	Cassie Blue A shared research platform where HZD-style controllers could be tested, compared, and extended.

The contribution of Hybrid Zero Dynamics was not simply that it made robots walk. It gave underactuated walking a mathematical design language. Dynamic legged locomotion involves continuous motion, impacts, changing contacts, underactuation, and stability questions that do not fit neatly into linear time-invariant control. HZD made it possible to design virtual constraints, induce a lower-dimensional invariant dynamics, and analyze periodic motion with tools such as Poincare maps while still respecting the hybrid nature of walking.

That framework mattered because it connected three worlds that are often separated: nonlinear control theory, numerical optimization, and hardware experiments. RABBIT showed that the ideas could survive contact with a physical underactuated robot. MABEL added compliance and dynamic running. MARLO moved toward 3D outdoor walking. Cassie Blue brought the methods onto a platform that many researchers could recognize, reproduce, and compare against.

The later rise of optimization, model predictive control, reinforcement learning, and data-driven locomotion did not replace this earlier work so much as change the surrounding toolchain. Classical nonlinear control supplied structure: models, stability questions, phase variables, virtual constraints, reduced dynamics, and experimentally meaningful benchmarks. Learning-based methods then entered a field where researchers already had better ways to ask what a gait was, how to test it, and how to move from equations to hardware.

In that sense, HZD belongs in the prehistory of modern learning-based legged locomotion. It did not cause the reinforcement-learning revolution, and it did not solve all of locomotion. But it helped make dynamic walking a rigorous engineering problem rather than a collection of impressive demonstrations. It showed that underactuation and impact could be embraced, not hidden, and that mathematical control design could produce controllers that worked on real robots.

This line of work also turned papers into a research ecosystem. The 2007 book Feedback Control of Dynamic Bipedal Robot Locomotion organized hybrid models, periodic orbits, virtual constraints, feedback design, and experimental implementation into a form that students and researchers could learn from. Later tools such as FROST and C-FROST helped shorten the path from full-order models to optimized gaits. Students and collaborators carried the methods into new laboratories, newer robot platforms, perception and state estimation, humanoids, exoskeletons, and learning-enhanced locomotion.

The translational thread is also real, though it should be stated with care. Work on lower-limb exoskeletons, including early control work connected to Wandercraft, drew on ideas developed for dynamic bipedal robots. The goal there was not a laboratory demonstration for its own sake, but self-balancing, hands-free locomotion in rehabilitation settings, including for people with paraplegia.

The durable legacy is therefore not a single robot or a single theorem. It is a research pathway: from geometric nonlinear control to hybrid zero dynamics, from analysis to optimization, from simulation to hardware, and from model-based control to the learning-rich methods now shaping legged robotics.