StarVLA QwenFAST for Bridge + RT-1 (10K)
This repository contains the uploaded 10,000-step checkpoint from the
1003_qwenfast run. It is configured as a StarVLA QwenFast policy trained on
the bridge_rt_1 mixture and evaluated on the four SimplerEnv WidowX tasks
whose logs are published beside the checkpoint. QwenFast encodes continuous
action chunks with the FAST tokenizer and trains the VLM by autoregressive
next-token prediction; it is not a DiT or diffusion policy.
Model details
| Item | Published configuration |
|---|
| Framework | StarVLA QwenFast |
| VLM | Local snapshot named Qwen2.5-VL-3B-Action; revision not recorded |
| Action encoding | FAST tokenizer over continuous action chunks |
| Training objective | Autoregressive cross-entropy over mapped action tokens |
| Inference | VLM token generation followed by FAST-token decoding |
| Action dimension / representation | 7 / delta_ee |
| Numeric policy-state input | None in the checked QwenFast forward and inference paths |
| Action chunk | 16 steps |
| Camera input | image_0, 224 x 224 |
| Dataset mixture | bridge_rt_1 (Bridge + RT-1/Fractal) |
| Normalization keys | bridge_dataset, fractal20220817_data |
| Uploaded checkpoint | checkpoints/steps_10000_pytorch_model.pt |
The YAML retains legacy dino, layer_qformer, and action_model_type: DiT-B
fields, but the selected QwenFast implementation does not construct those
components: it uses the FAST tokenizer and the VLM language-model loss. The
action-model section does not declare state_dim, while the packaged raw
statistics contain an 8D state for both datasets. Those statistics are a
dataset artifact, not evidence that a numeric state vector enters this policy.
Training details
| Setting | Value in config.yaml |
|---|
| Intended maximum steps | 100,000 |
| Released and evaluated step | 10,000 |
| 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 / 0.1 |
Packaged freeze_modules | Boolean true; the public trainer only accepts a comma-separated string, so this does not select modules to freeze |
| Seed | 42 |
The YAML records the run target, not the released training extent. No weight
later than 10K is present in this repository.
SimplerEnv WidowX evaluation
The repository includes four run logs for each of four WidowX tasks. Each
completed log reports the standard 24-episode SimplerEnv success average. The
table below is the unweighted mean of the four published run-level
Average success values for this exact 10K checkpoint.
| Task | Success rate |
|---|
| PutCarrotOnPlateInScene-v0 | 0.4167 |
| PutEggplantInBasketScene-v0 | 0.8438 |
| PutSpoonOnTableClothInScene-v0 | 0.7188 |
| StackGreenCubeOnYellowCubeBakedTexInScene-v0 | 0.3646 |
| Mean over all 16 run logs | 0.5859 |
The resulting 58.6% rounded value matches the entry for this repository in
the
StarVLA Model Zoo.
These are simulation results; they do not establish real-robot performance.
Files
1config.yaml
2dataset_statistics.json
3checkpoints/
4├── steps_10000_pytorch_model.pt
5└── steps_10000_pytorch_model_infer_*.log.run{1,2,3,4}
Loading and evaluation
1huggingface-cli download StarVLA/Qwen-FAST-Bridge-RT-1 \
2 --local-dir Qwen-FAST-Bridge-RT-1
3
4export CKPT=$PWD/Qwen-FAST-Bridge-RT-1/checkpoints/steps_10000_pytorch_model.pt
5python deployment/model_server/server_policy.py \
6 --ckpt_path "$CKPT" \
7 --port 10093 \
8 --use_bf16
Use the maintained
SimplerEnv integration
for the client. Select
bridge_dataset for WidowX evaluation; the second
statistics key belongs to the RT-1/Fractal component of the training mixture.
Intended use and limitations
This checkpoint is intended for manipulation-policy research and SimplerEnv
WidowX evaluation with a single RGB view and 7D delta-end-effector actions. The
exact upstream VLM revision is absent from the Hub config. It is not a
Transformers model directory, has not been shown safe
on physical hardware, and is not safety-tuned. Loading requires the matching
action-token vocabulary and FAST tokenizer; treating the stale DiT-B YAML
label as the runtime decoder would reconstruct the wrong model.
