THETA Bench GR00T N1.7: Real G1 additional training, 91 demonstrations

The validated final policy checkpoint is available in this repository.

This is a THETA training-result repository. It does not substitute an upstream pretrained policy for a THETA-trained checkpoint.

Setting	Value
Training stage	Real G1 additional training, 91 demonstrations
Target optimizer updates	5000
Per-GPU batch / GPUs / global batch	16 / 8 / 128
Gradient accumulation	1
Conditions per global batch	4
Dataset revision	47eca9322bb53fa1c685363271a87d2e414cb0e8

Pinned training data.

Initialization: snupilab/theta-bench-groot-sim-3003 after 40,000 simulation updates, followed by 5,000 new updates using the 91 real G1 demonstrations.

The four real conditions are StickMove Standard/Reasoning and HookRetrieve Standard/Reasoning. Recordings are 20 Hz. Hardware executed joint-target actions require the matching real G1 control adapter. A simulation adapter must not be assumed compatible.

Use the model's native THETA adapter and model-specific dependencies. This repository does not claim compatibility with arbitrary Transformers or simulation loaders. No evaluation score is claimed by checkpoint publication.

This final checkpoint has 32 state dimensions and 35 supervised real G1 action dimensions at 20 Hz, with a 40-frame horizon. Load it through native Gr00tPolicy using its saved real modalities and matching hardware executed-joint-target adapter. The current THETA simulation _groot_server enforces 36-dimensional simulation actions and is incompatible with this real checkpoint. The native constructor requires backbone weights and processor from nvidia/Cosmos-Reason2-2B at revision 9ce19a195e423419c349abfc86fd07178b230561 before restoring the trained state. Keep that dependency available separately. Native fine-tuning initialization was the corresponding validated THETA simulation checkpoint at update 40000.

Training uses independent model optimizers and shared GPU execution through MPS. Publication is performed by a CPU uploader after final checkpoint validation.