A Unitree G1 humanoid completed all 30 locomotion and posture trials using VioLA, which maps human and robot body and hand movement into shared motion codes that controllers turn into joint and finger commands. The study was posted Oct. 8. Human recordings made up 93.2% of the 140.6 million-frame training pool. The tests required no task-specific fine-tuning, although those task families appeared in training.
Robot policies usually learn outputs such as joint movements, but a human video does not contain commands for a robot’s motors. VioLA uses pretrained encoders to map human and robot demonstrations into matching body and hand motion spaces, giving one generalist policy a common target to predict.
The main policy trained first on human demonstrations, then on a mixture of human and robot data. On a physical Unitree G1, it completed five trials each for six tasks such as walking, bending and standing up from a chair. The released GR00T N1.7 model completed 5 of those 30 trials, while Ψ0 completed none. In seven object-handling tasks, VioLA completed 31 of 35 trials, or 88.6%.
A separate version trained the generalist policy with human demonstrations alone completed 12 of 30 locomotion trials. That result still relies on a pretrained backbone and motion modules; “human-only” describes the demonstrations used to train the policy, not a system built from scratch.
The authors’ “zero-shot” label means that VioLA received no task-specific fine-tuning. The task families were in its training data, while the particular test episodes, laboratory setting and object instances were held out. The evaluation covered 13 main tasks on one robot, with five trials apiece. A separate bottle-carrying test failed in all five attempts; counting it lowers the manipulation result to 31 of 40, or 77.5%. The system also depends on annotated human motion and a hand decoder that has no contact-force feedback.
Sources
- Albaba et al., “VioLA: Learning Generalist Humanoid Control Policies from Human Data,” arXiv:2610.12435v1
- Authors’ VioLA project page and results
- Illustrative photo, not from the study: Unitree G1 at UAV Expo 2024, by Sayanesy; CC0 1.0.