Release the model after UMI trained (before the real robot sft)

[lucasjinreal]

How much does the UMI data trained gains compare with directly sft on real robot data?
How does the UMI data being used to train the network? How does the action represent?

[cghezhang]

Q1: How does training with UMI data compare to directly fine-tuning on real robot data?
We found that VLA pre-training on 5K hours of UMI data leads to a clear improvement in downstream task accuracy, which is also reflected in the SFT loss curves.

Q2: How is the UMI data used for training?
We use 5K hours of UMI data for VLA pre-training before task-specific SFT. The model is trained for 200K iterations with a total batch size of 1024.

Q3: How are the actions represented?
Actions are represented as relative end-effector poses with respect to the first frame of each chunk.

Thank you for your interest! We will release a report on our VLA model and dataset soon with more details！

[lucasjinreal]

Hi, thanks for the detailed messages.

Still wanna ask: 1. does your VLA also using flowmatching? If so, does the lm head and NTP loss applied or totally removed? 2. does the action relative as well for rotation and gripper