Qwen3-VL-4B GR00T for Bridge + RT-1 (20K)
This repository contains a StarVLA QwenGR00T WidowX policy trained on the
Bridge and Fractal/RT-1 LeRobot datasets. The action decoder is a diffusion
transformer flow-matching head; this is not an OFT regression model.
Checkpoint identity
| Item | Value |
|---|
| Released file | checkpoints/steps_20000_pytorch_model.pt |
| Training step | 20,000 |
| Hub revision checked | 12acc0b0f1f6230df21c479934a67a930b52f878 |
| File size | 9,976,845,210 bytes |
| SHA-256 / LFS object ID | 769d6c400d582a86ae8df8b0b445240ab679dbe77eeb72a4db71e43cd129c7c3 |
Although the configured run target is 100K and the summary contains later save
events, no checkpoint later than 20K is published here.
Model and control contract
| Item | Value |
|---|
| Framework | StarVLA QwenGR00T |
| Base VLM | Qwen3-VL-4B-Instruct |
| Action model | 16-layer DiT-B flow head: 768 latent width, 12 heads (64 dimensions/head); state/action decoder MLP width 1,024 |
| Observation | Language + one image_0 RGB view at 224 x 224 |
| Policy action / state dimensions | 7 / 7; packaged dataset state statistics contain 8 values |
| Action chunk | 16 x 7 |
| Action representation | delta end-effector action (delta_ee) |
| Inference integration | 4 flow steps |
| SimplerEnv normalization | oxe_bridge for WidowX Bridge evaluation |
| Action ensemble horizon | 7 |
The SimplerEnv adapter sends image and language without robot state. Do not
silently change the observation payload, action coordinate system, or
normalization key. The YAML also carries a dinov2_vits14 field, but the
matched QwenGR00T implementation does not construct a DINO or Q-Former module.
Training data and settings
The bridge_rt_1 mixture combines the packaged oxe_bridge statistics
(1,284,318 transitions, 53,192 trajectories) and oxe_rt1 statistics
(3,449,678 transitions, 87,212 trajectories).
| Setting | Value |
|---|
| Per-device VLA batch size | 16 |
| Gradient accumulation | 1 |
| Training flow repeats | 4; the public run-era framework reads trainer value 4, not the YAML action-model value 8 |
| Optimizer | AdamW, betas (0.9, 0.95), epsilon 1e-8, weight decay 1e-8 |
| Warmup | 10,000 steps |
| VLM co-training loss scale | 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 |
| Training GPU count | Missing from the public artifact |
Packaged-log evaluation
The repository packages 16 logs whose filenames and final command lines point
to steps_20000_pytorch_model.pt: four SimplerEnv Bridge tasks, four runs per
task, and 24 episodes per run (384 episodes total).
| Task | Success |
|---|
| Put carrot on plate | 59.38% |
| Put eggplant in basket | 100.00% |
| Put spoon on tablecloth | 75.00% |
| Stack green cube on yellow cube | 18.75% |
| Micro-average, 384 episodes | 63.28% |
These values are recomputed from the packaged logs. The StarVLA model-zoo page
lists 65.3%, but that number is not reproduced by this artifact's 16 logs and
is therefore not used as the checkpoint score here.
Download and load
1hf download StarVLA/Qwen3VL-GR00T-Bridge-RT-1 \
2 --local-dir playground/Pretrained_models/Qwen3VL-GR00T-Bridge-RT-1
3
4export CKPT=playground/Pretrained_models/Qwen3VL-GR00T-Bridge-RT-1/checkpoints/steps_20000_pytorch_model.pt
5python deployment/model_server/server_policy.py \
6 --ckpt_path "$CKPT" \
7 --config_override framework.qwenvl.base_vlm=Qwen/Qwen3-VL-4B-Instruct \
8 --port 57700 \
9 --use_bf16
Evaluate through the WidowX Bridge adapter documented in
examples/simBenchmarks/SimplerEnv/README.md, using oxe_bridge, delta-EEF
actions, and the same ensemble settings.
Intended use and limitations
The policy is intended for the four WidowX Bridge tasks in SimplerEnv. Its
results do not establish performance on RT-1 robot hardware, other embodiments,
or real-world deployment. Preserve the matching StarVLA framework and treat
the pickle-based checkpoint as executable data from a trusted revision.
License status
This target repository did not previously publish a Model Card or a separate
LICENSE file. The checkpoint's weight license therefore needs maintainer
confirmation; the Qwen3-VL base-model and training-dataset terms still apply.
