StarVLA/Qwen-GR00T-Bridge-RT-1

StarVLA legacy QwenFM checkpoint for Bridge + RT-1 (30K)

This repository is named Qwen-GR00T-Bridge-RT-1, but its published config.yaml selects framework_py: QwenFM and a local VLM path named nora. The card preserves that discrepancy instead of relabeling the checkpoint as a verified current QwenGR00T artifact.

Model details

Item	Published configuration
Hub repository name	Qwen-GR00T-Bridge-RT-1
Configured framework	Legacy StarVLA QwenFM
Configured VLM	Local snapshot ./playground/Pretrained_models/nora; identity/revision missing
YAML-only visual field	dinov2_vits14; the matched legacy QwenFM implementation does not construct DINO or Q-Former
Action model	16-layer DiT-B flow head: 768 latent width, 12 heads (64 dimensions/head); state/action decoder MLP width 1,024
Action / model-state dimension	7 / 7
Raw state statistics	8D for both oxe_bridge and oxe_rt1
Action representation / chunk	delta_ee / 16 steps
Inference flow steps	4
Camera input	image_0, 224 x 224
Uploaded checkpoint	checkpoints/steps_30000_pytorch_model.pt

Because the VLM is identified only by a local directory name, no Hugging Face base_model is asserted in the metadata. Reconstruct the checkpoint using the packaged YAML and the compatible StarVLA revision; do not substitute a current backbone solely from this repository's display name.

Training details

Setting	Value in config.yaml
Dataset mixture	bridge_rt_1
Intended maximum steps	100,000
Released and evaluated step	30,000
Per-device VLA batch size	16
Gradient accumulation	1
Warm-up steps	10,000
Base / interface / action LR	3e-5 / 1e-5 / 1e-4
Optimizer	AdamW, betas (0.9, 0.95), epsilon 1e-8
Scheduler	Cosine with minimum LR 5e-7
VLA / VLM loss scale	1.0 / 0.05
freeze_modules	Packaged boolean true; the public trainer expects module paths as a string, so this value names/selects no modules
Seed	42

SimplerEnv WidowX evaluation

The repository publishes four completed run logs for each WidowX task. The values below are recomputed from their terminal Average success fields.

Task	Success rate
PutCarrotOnPlateInScene-v0	0.5312
PutEggplantInBasketScene-v0	0.5938
PutSpoonOnTableClothInScene-v0	0.9167
StackGreenCubeOnYellowCubeBakedTexInScene-v0	0.4375
Mean over all 16 run logs	0.6198

This value is tied to the uploaded logs. The current Model Zoo has historical PI/GR00T naming and link inconsistencies for the Qwen2.5 Bridge + RT-1 rows, so its 63.6 entry is not attached to this artifact here.

Files

config.yaml
dataset_statistics.json
checkpoints/
├── steps_30000_pytorch_model.pt
└── steps_30000_pytorch_model_infer_*.log.run{1,2,3,4}

Loading and evaluation

huggingface-cli download StarVLA/Qwen-GR00T-Bridge-RT-1 \
  --local-dir Qwen-GR00T-Bridge-RT-1

export CKPT=$PWD/Qwen-GR00T-Bridge-RT-1/checkpoints/steps_30000_pytorch_model.pt
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" \
  --port 10093 \
  --use_bf16

For WidowX evaluation use the SimplerEnv client and normalization key oxe_bridge. oxe_rt1 is the second training-mixture key.

Intended use and limitations

This is a legacy research artifact with an unresolved public identity mismatch between the Hub name, configured framework, and local VLM path. Its bundled SimplerEnv logs are usable as evidence for the 30K checkpoint, but the artifact should not be advertised as a drop-in current QwenGR00T model until the framework/backbone provenance is confirmed. It is not safety-tuned.