A research paper teaches a humanoid robot from AI-generated motion, without real-world demonstrations

Teaching humanoid robots new tasks usually means strapping people into motion-capture suits and recording them perform each task, often multiple ways. A new arXiv preprint (Learning Diverse Humanoid Tasks via Synthetic Video Scenarios without Real World Data) proposes skipping that step: type a prompt, let a generative video model produce several versions of a human performing the same action, and train a Unitree G1 on the synthetic clips alone.

A text prompt goes to Google Veo 3.1, which outputs realistic videos of human movement. Those frames are run through SMPL-X, a public body model, to extract skeletal poses. GMR (Generic Motion Retargeting) and inverse-kinematics optimization reshape the motion for the robot, and a reinforcement-learning policy with PPO-style updates and a privileged critic tracks it. The reward combines pose, velocity, end-effector, and center-of-mass terms.

The authors motivate the work against the conventional approach of learning from human demonstrations, where capture cost, per-individual variation, and the fact that humans do the same task differently each time all make scaling hard. Their counter: a generative model can synthesize the variation for them.

The evidence ceiling is explicit. The paper (HTML) reports results only in four Unitree G1 simulation scenarios, with no real-robot transfer, no out-of-distribution test, and no comparison to existing demonstration pipelines. It is also a preprint, not yet peer-reviewed. The pipeline is named; the four-scenario simulation result is the entire empirical claim.