Physical AI Lab // Episode 3.5 // One-pager

GR00T fine-tuning in practice

Tuning an open humanoid model to your own robot

Three key points

  1. GR00T N1.7 is an open, commercially usable humanoid VLA: 3 billion parameters, a dual-system architecture (O), pre-trained on 20,854 hours of human egocentric video (O). A solid foundation.
  2. Fine-tuning starts with data: put it in the LeRobot format and write the modality config (wiring each dimension to a joint). Practise on a small example dataset first; which conversion scripts and examples the repo offers follows the repo's instructions.
  3. One command trains a little more, on a single GPU (GPU and memory requirements per the repo's instructions). Check it with an open-loop evaluation (predicted vs real actions), but good open-loop results don't mean a stable closed loop on a real robot.

Evidence: O official · R research · M media

O3 billionparameters: GR00T N1.7 is an open 3 billion parameter vision-language-action model (sees images, takes instructions, outputs continuous actions)
OCommercial uselicence: NVIDIA describes GR00T N1.7 as open and commercially licensed
O20,854 hoursof human egocentric video used to pre-train GR00T N1.7 (EgoScale, over 20 task categories)
OMore than doubleNVIDIA says: going from 1,000 to 20,000 hours of human video more than doubles average task completion (dexterity scaling law)
O22 DoFhands: this dexterity law lets 22 degree-of-freedom dexterous hands do contact-rich fine tasks

Further reading