snupilab/humanoidtoolbench-dp-real-g1-91

HumanoidToolBench DP: Real G1 additional training, 91 demonstrations

The validated final policy checkpoint is available in this repository.

This is a HumanoidToolBench training-result repository. It does not substitute an upstream pretrained policy for a HumanoidToolBench-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/humanoidtoolbench-dp-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 HumanoidToolBench 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.

Retain run/checkpoints/ckpt_5000 layout. Use the HumanoidToolBench native ACT/DP adapter with the exported run directory and final step. Run from that directory so dataset_statistics.json resolves. Preserve launch.json's exact language receipt. Install pinned Psi0/HumanoidToolBench source and pre-cache openai/clip-vit-large-patch14@32bd64288804d66eefd0ccbe215aa642df71cc41. DP inference uses ema_net.pth; model.safetensors is retained for native checkpoint discovery. Real G1 uses executed joint targets at 20 Hz with the absent final yaw unsupervised; hardware deployment is not qualified by this export.

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