Request for real-robot pretraining dataset_stats.json
Thank you for releasing the pretrained weights.
I am trying to deploy and fine-tune the model on real robots.
Since using incorrect stats from other datasets will lead to unsafe and unexpected robot movements, could you please share the dataset_stats.json (or normalization statistics) used during your real-robot pretraining?
This would greatly help the community proceed with proper downstream fine-tuning. Thanks!
Thank you for your interest, and apologies for the inconvenience!
Unfortunately, due to the time that has passed since the pretraining, we can no longer locate the dataset_stats.json file used for RoboCoin's real-robot pretraining. This was an oversight on our part, and we sincerely apologize for not preserving these statistics.
That said, we would like to suggest an alternative approach for downstream fine-tuning.
Since the amount of data used for RoboCoin pretraining was relatively limited, we do not expect the pretrained model to achieve strong zero-shot, out-of-the-box performance on new real-world robot setups. Therefore, when performing post-training based on the AHA-WAM-Pretrained checkpoint, we believe it should be reasonable to directly use the dataset_stats.json (normalization statistics) computed from your own post-training dataset, rather than trying to reproduce the original pretraining statistics.
In particular, we expect the pretrained representations to remain useful as long as the downstream setup is reasonably aligned with the original training setup in the following aspects:
- Robot configuration: A robotic arm with a configuration broadly similar to the AgileX Piper.
- Action space: Using joint-space actions (joint positions) rather than a fundamentally different action representation.
Under these conditions, we believe the model should still be able to leverage a substantial portion of the pretrained knowledge, even when using normalization statistics from your own dataset.
We hope this provides a useful path forward, and thank you again for your interest in our work!