Availability of G1 post-training datasets & eval sets (+ question on data scale for new-scene generalization)

[Gumin-Research]

Hi, thanks for open-sourcing DreamDojo - the pretrained action-conditioning transfers remarkably well.

We're post-training the 2B G1 model on our own Unitree G1 teleoperation data and seeing strong results: within ~1k steps the model adapts to our scenes, renders previously-OOD tasks coherently, and AR drift over multi-chunk rollouts is greatly reduced.

1. Data availability

The G1 experiment configs (e.g. configs/2b_480_640_g1.yaml) point dataset_path to internal paths that aren't downloadable, e.g.:

/mnt/amlfs-02/shared/datasets/standardized_v1/g1/unitree_g1.g1-shelf_corrected-waist_aug18-sep3
/mnt/amlfs-02/shared/datasets/standardized_v1/g1/unitree_g1.in-the-wild_nov2-nov12
... (corl-demo / g1-cart / g1-corl-basket / in-the-wild campaigns)

The README notes the GR-1 post-training datasets and evaluation sets are released(PhysicalAI-Robotics-GR00T-Teleop-GR1), but we don't see the G1 equivalents.

• Are the G1 post-training datasets (shelf / basket / cart / in-the-wild) available or planned for release?
• Are the G1 evaluation sets (In-lab / EgoDex / DreamDojo-HV for G1) available?

Having these would let us reproduce the reported G1 object/environment generalization and benchmark our own post-training against the official recipe.

2. Question on data scale for new-scene generalization

On scenes/tasks within our post-training distribution the results are great. On a genuinely new scene/task (a different lab, an object-interaction not in our data) we observe a consistent pattern: the environment/background still renders, but the robot arm degrades under AR drift and the rollout collapses over longer horizons.

This holds across checkpoints (10k and 50k behave similarly), so it seems to be a coverage/OOD effect rather than overfitting.

Given the paper's claim that a target dataset "can be collected in limited domains at a small scale," could you share any guidance on:

• Roughly how much / how diverse target-robot data was used per embodiment for the reported generalization?
• Whether new-scene generalization (beyond the post-training scenes) is expected from the released checkpoints, or requires the scene to be represented in post-training data?

This is partly why we're asking about the G1 datasets above - we'd like to understand the data scale needed to cover new environments.

Thanks!