Nvidia GR00T: Robot training accelerates with expert intervention

TL;DR. Nvidia GR00T model training dramatically improved by integrating expert-corrected failure episodes into its dataset. - Robotics team used teleoperation to fix autonomous inference failures, converting these interventions into finetuning data. - Adding just 100 expert episodes boosted the success rate from 62% to 93%, a significant 31 percentage point gain. - This method allows robots to learn from errors more efficiently than collecting only successful run data.

• Expert intervention through teleoperation significantly improves VLA model performance on edge cases.
• Correcting autonomous inference failures and using the recovery data for fine-tuning is more efficient than collecting new successful runs.
• A small dataset of 100 expert intervention episodes led to a 31 percentage point increase in robot success rate for the NVIDIA GR00T model.
• This approach provides a viable alternative for addressing unpredictable robot behaviors without restarting training from scratch.

Back to QLANKR News