See HumanoidOS In Action
- The Level 1 pre-trained model has seen over a year's worth of data. The fine-tuned model has seen an additional 100,000 trajectories specific to the embodiment.
- The pre-trained model gives the robot general skills, and fine-tuning trains the diffusion transformer and embodiment
specific encoders-decoders to perform motions with human-level precision
- The pre-trained model is trained at each level with an amount of data double to what was used in the previous stage. The fine-tuning dataset is increased in the same fashion, enabling continuous improvement over time.