How do we solve unpredictable edge cases in VLA models?
In our latest experiments with the NVIDIA GR00T model, our robotics team tackled this exact challenge by leveraging expert intervention.
When autonomous inference failed, we did not restart training from scratch. Instead, we used teleoperation to immediately correct the action and converted this recovery process into finetuning data.
Adding just 100 of these expert intervention episodes to our initial 150 episodes boosted our success rate from 62 percent to 93 percent, showing a 31 percentage point improvement. While this might not be the absolute best solution for every scenario, we believe it could be a highly viable alternative for addressing edge cases.
This demonstrates that teaching robots how to recover from failures is far more efficient than simply collecting more successful run data.
We are sharing our test video and technical manual below. Feel free to check them out.
Video: https://youtu.be/CRoOspJ91qk
Open Source (ROS 2): GitHub - ROBOTIS-GIT/cyclo_intelligence: cyclo_intelligence · GitHub
Docs: Cyclo Intelligence | ROBOTIS