Last updated: 2026-11-04
Humanoid parkour is not new. Boston Dynamics’ hydraulic Atlas was vaulting and backflipping years ago. Those routines, though, relied on carefully planned sequences and hand-built controllers. The research question since then has been whether a robot can pick and chain such moves itself, from what it sees, in places it has never been.
Lightbot 0 is Light Origins’ answer for a small, cheap body. One neural-network policy, running on the robot, takes in a depth image, the robot’s own body state and a direction to go, and decides moment by moment whether to keep walking, climb, vault or step down. Nothing external tells it which skill to use.
The training recipe starts with a short video of a person doing a move. Light Origins extracts the timing and the points of contact, rebuilds the motion inside a physics simulator within the robot’s actuator limits, and then makes the task harder step by step, raising the obstacle 5–10 cm at a time from a 45 cm start. Specialist teacher policies that can “see” perfect terrain data are trained first; their knowledge is then distilled into one student that works from the depth camera alone.
A memory unit in the student lets it remember an obstacle after it drops out of view, which matters mid-vault when the camera points away. In Light Origins’ simulation tests, removing that memory cut success at 70 cm from 90% to 34%.
The hardware demos are impressive for a 90 cm robot: obstacles approaching the robot’s own shoulder height, narrow beams and stairs, with skills transferring from simulation without real-world fine-tuning. But every success rate Light Origins has published comes from simulation. The real-robot results are shown, not counted. The limits are visible too: at 75 cm the deployable policy succeeded only a third of the time in simulation.
The method needs little expensive infrastructure. No motion-capture rig, no pre-built maps and no offboard computer are required, only short human clips and simulation time. If that recipe scales to larger robots, adding a new physical skill to a humanoid could look more like collecting a few examples than writing a new controller.
See full timeline →