StarVLA/Qwen-GR00T-Bridge

StarVLA QwenGR00T for Bridge (45K)

This repository contains the uploaded 45,000-step checkpoint from the 1019_qwen25_groot_bridge_only run. Unlike the similarly named Bridge + RT-1 repositories, its data_mix is bridge, so the published training statistics contain only the Bridge/WidowX embodiment.

Model details

Item	Published configuration
Framework	StarVLA QwenGR00T
Base VLM	StarVLA/Qwen2.5-VL-3B-Instruct-Action
YAML-only visual field	dinov2_vits14; the matched QwenGR00T implementation does not construct DINO or Q-Former
Action model	16-layer DiT-L flow head: 1,536 latent width, 32 heads (48 dimensions/head); state/action decoder MLP width 1,024
Action / model-state dimension	7 / 7
Raw state statistics	8D under oxe_bridge
Action representation / chunk	delta_ee / 16 steps
Inference flow steps	4
Camera input	image_0, 224 x 224
Uploaded checkpoint	checkpoints/steps_45000_pytorch_model.pt

The 8D raw state statistics and 7D action-head state are reconciled by the packaged data transform. Do not bypass that transform by feeding a guessed flat vector. The YAML's DINO field is inherited configuration, not a runtime component of this QwenGR00T path.

Training details

Setting	Value in config.yaml
Intended maximum steps	100,000
Released and evaluated step	45,000
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
Training flow repeats / inference steps	4 / 4; the public run-era framework reads trainer value 4, not the YAML action-model value 8
VLA / VLM loss scale	1.0 / 0.1
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

Four run logs per task are bundled with the 45K checkpoint. The following values are means of the four published run-level success rates for each task.

Task	Success rate
PutCarrotOnPlateInScene-v0	0.5625
PutEggplantInBasketScene-v0	0.8750
PutSpoonOnTableClothInScene-v0	0.9375
StackGreenCubeOnYellowCubeBakedTexInScene-v0	0.4792
Mean over all 16 run logs	0.7135

This artifact-derived 71.35% rounds to the 71.4 value published for the same repository in the StarVLA Model Zoo.

Files

config.yaml
dataset_statistics.json
checkpoints/
├── steps_45000_pytorch_model.pt
└── steps_45000_pytorch_model_infer_*.log.run{1,2,3,4}

Loading and evaluation

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

export CKPT=$PWD/Qwen-GR00T-Bridge/checkpoints/steps_45000_pytorch_model.pt
python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" \
  --port 10093 \
  --use_bf16

Use the StarVLA SimplerEnv evaluator with normalization key oxe_bridge.

Intended use and limitations

The model is scoped to Bridge/WidowX simulation with one RGB camera and 7D delta-end-effector control. It was not trained on RT-1 according to the packaged YAML, despite nearby similarly named releases. Real-robot performance, safety, and transfer to other observation/action mappings are not established.