Lightbot 0

Lightbot 0 is a compact humanoid robot built by Light Origins, a robotics company with offices in Beijing, Shenzhen and Singapore. It stands 90 cm tall, weighs 18.9 kg and has 21 actuated joints. In August 2026 the team used it to show a new way of teaching humanoids parkour: speed vaults, reverse vaults, climbing onto boxes and stepping down, plus stairs, a narrow plank bridge and stepping stones.

What makes the demo notable is how little the robot relies on. A single recurrent neural-network policy runs on the robot at 50 times a second. Its only inputs are a chest-mounted depth camera, its own joint and body sensing, and a velocity command. There is no motion capture, no pre-loaded map, no skill labels and no scripted switching between moves; the policy itself decides when to walk, climb, vault or step down. When an obstacle drops out of the camera’s view mid-vault, the robot relies on a short-term memory of what it saw.

The training method, which Light Origins calls Real2Sim2Real (LightParkour), starts from a short clip of a human performing a move. The motion is adapted to the robot and repaired in physics simulation, then pushed to harder versions step by step, raising obstacles 5–10 cm at a time from a 45 cm starting point. Four specialist “teacher” policies are then distilled into one student that runs on depth images alone. The skills transferred to the real robot zero-shot, with no real-world fine-tuning.

Read the numbers carefully. All published success rates come from simulation: 90% for climbing onto a 70 cm box (77% of the robot’s height) across 500 randomised trials, and 98% for chained skills with learned transitions, against 33% for isolated skills. At 75 cm the deployable policy managed 33.4%. Real-world performance is shown in demos but not measured. Lightbot 0 is a research platform with no announced price or availability.