THG News · Research

Berkeley's Cassie+ Learns to Walk Sideways Without Human Demonstrations

June 21, 2026

Sergey Levine's lab at UC Berkeley published results on June 21 showing the Cassie+ bipedal robot learning previously unseen locomotion gaits — including sideways gallop and one-legged hopping — purely from self-supervised reinforcement learning, with no human demonstrations.

The Technique

The team's "Curiosity-Driven Locomotion" framework uses an intrinsic reward derived from prediction error in a learned world model. The robot is incentivized to discover its own novel, stable gaits.

Why It Matters

Most current humanoid locomotion policies are bootstrapped from human motion-capture data, which limits the gait library to what humans can demonstrate. The Berkeley work suggests that robots can — and should — discover gait modalities optimized for their own kinematics rather than imitating human walking.

The work has obvious implications for non-anthropomorphic embodiments and is being closely watched by Boston Dynamics and Agility Robotics, both of which have publicly cited "embodiment-native locomotion" as a 2027 priority.