StarVLA QwenGR00T Qwen3-VL-4B for RoboCasa GR1 (90K)
This repository contains the uploaded 90,000-step checkpoint from the
1226_fourier_gr1_unified_1000_QwenGR00T_qwen3 run. It is a state-conditioned
StarVLA QwenGR00T policy trained on the 24-task Fourier GR1 RoboCasa mixture.
Although the run configuration targets 100K and summary.jsonl lists a 100K
step, the only published weight file is the 90K checkpoint.
Model details
| Item | Published configuration |
|---|
| Framework | StarVLA QwenGR00T |
| VLM | Local snapshot named Qwen3-VL-4B-Instruct-Action; revision not recorded |
| Action model | 16-layer DiT-B flow head: 768 latent width, 12 heads (64 dimensions/head); state/action decoder MLP width 1,024 |
| Action dimension | 29 |
| Raw / model state dimension | 29 / 58 |
| Action chunk | 16 steps |
| Inference flow steps | 4 |
| Camera | Single video.ego_view, 224 x 224 |
| Dataset mixture | fourier_gr1_unified_1000 |
| Normalization key | gr1 |
| Uploaded checkpoint | checkpoints/steps_90000_pytorch_model.pt |
The public GR1 data configuration orders state and action as left arm (7),
right arm (7), left hand (6), right hand (6), and waist (3), for 29 raw
dimensions. StateActionSinCosTransform expands the raw state to the model's
configured 58D representation. Actions use per-key min-max normalization.
Training data
The mixture contains 24 GR1 tabletop pick-and-place tasks, with 1,000
trajectories named for each task in the public registry. The packaged
statistics report 24,000 trajectories under the gr1 embodiment key.
Training details
| Setting | Value in config.yaml |
|---|
| Intended maximum steps | 100,000 |
| Released step | 90,000 |
| Resume mode | true |
| Per-device VLA batch size | 8 |
| Gradient accumulation | 1 |
| Warm-up steps | 5,000 |
| Base / interface / action LR | 3e-5 / 1e-5 / 1e-4 |
| Optimizer fields | Betas (0.9, 0.95), epsilon 1e-8, weight decay 1e-8; name not stored |
| Scheduler | Cosine with minimum LR 5e-7 |
| Diffusion repeats / inference steps | 4 / 4 |
freeze_modules | Packaged boolean true; the public trainer expects module paths as a string, so this value names/selects no modules |
| Seed | 42 |
No training hardware, global batch size, precision mode, or exact resume source
is recorded in the Hub YAML, so those fields are not inferred here.
Related project-reported evaluation
The public
RoboCasa GR1 evaluation guide
links this repository for
Qwen3VL-GR00T and reports the following result:
| Reported model | Tasks | Rollouts per task | Average success rate |
|---|
StarVLA-GR00T-Qwen3 | 24 | 50 | 47.8% |
The published table names the model family but does not record a checkpoint
step, SHA, or evaluation log bundle. Therefore 47.8% is a related
project-reported result and is not presented as independently verified for
the uploaded 90K artifact. Reproducible attribution requires the original
rollout outputs or a checkpoint identifier from the evaluation run.
Files
1config.yaml
2dataset_statistics.json
3summary.jsonl
4checkpoints/
5└── steps_90000_pytorch_model.pt
Keep the configuration and gr1 statistics with the checkpoint. The summary
file enumerates saved training steps but contains no evaluation metrics.
Loading and evaluation
1huggingface-cli download StarVLA/Qwen3-VL-GR00T-Robocasa-gr1 \
2 --local-dir Qwen3-VL-GR00T-Robocasa-gr1
3
4export CKPT=$PWD/Qwen3-VL-GR00T-Robocasa-gr1/checkpoints/steps_90000_pytorch_model.pt
5python deployment/model_server/server_policy.py \
6 --ckpt_path "$CKPT" \
7 --port 5678 \
8 --use_bf16
Run the RoboCasa simulator in a separate environment following the evaluation
guide linked above. Keep state input enabled for this GR00T checkpoint and use
normalization key gr1. The --no_send_state note in that guide applies to
the separate QwenOFT release, not this model.
Intended use and limitations
This checkpoint is intended for the 24-task RoboCasa GR1 tabletop simulation
with the exact ego-camera, state/action ordering, and normalization contract
above. The uploaded files do not prove the reported 47.8% result belongs to
step 90K, and they do not document the resume source or VLM revision. Transfer
to other GR1 stacks, camera layouts, robots, or real hardware is unverified.
The policy is not safety-tuned.
