Qwen3-VL-4B OFT for RoboTwin 2.0 Base (40K)
This repository contains a StarVLA QwenOFT checkpoint trained on the
RoboTwin 2.0 Base mixture: 50 clean demonstrations for each of 50 tasks. It is
a complete StarVLA state dictionary rather than a Transformers model package.
Checkpoint identity
| Item | Value |
|---|
| Released file | checkpoints/steps_40000_pytorch_model.pt |
| Training step | 40,000 |
| Hub revision checked | bf41eb7c9edb697a5b2677e719beb042bcfe9ccb |
| File size | 9,785,132,794 bytes |
| SHA-256 / LFS object ID | 8ccee8ddb33fa5df83a9a5e2b7d4f737db93f19a051b23f738af8a484fe2d7e5 |
The run history reaches 100K, but only the 40K weights are published here.
Model and control contract
| Item | Value |
|---|
| Framework | StarVLA QwenOFT |
| Base VLM | Qwen3-VL-4B-Instruct |
| Action head | Two-block residual MLP, 2,560 input / 5,120 hidden / 14 output; direct L1 regression |
| Camera order | head, left wrist, right wrist |
| Robot state | Statistics are packaged, but the current absolute-action evaluator omits state from the policy request |
| Action chunk | 16 x 14 |
| Action representation | 14-D absolute dual-arm joint action |
| Normalization | min-max with key new_embodiment |
Preserve the RoboTwin adapter's action reordering when converting the model's
14-D output to simulator control. A numerically correct tensor with a different
joint order is not an equivalent policy input or output.
Training data and settings
dataset_statistics.json records 2,500 trajectories and 549,787 transitions
for new_embodiment, matching 50 tasks x 50 clean demonstrations.
| Setting | Value |
|---|
| Dataset mixture | robotwin |
| Per-device VLA batch size | 8 |
| Gradient accumulation | 1 |
| Optimizer | AdamW, betas (0.9, 0.95), epsilon 1e-8, weight decay 1e-8 |
| Base / interface / action LR | 3e-5 / 1e-5 / 1e-4 |
| Warmup | 5,000 steps |
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 |
Repository-reported result boundary
The StarVLA RoboTwin README links this repository and reports 50.38% Easy
success for its Qwen3-VL-4B Base column. The page does not identify the
evaluated save step and this repository contains no raw evaluation logs.
Therefore, 50.38% is useful repository-level context but is not verified as a
score of the released 40K file. A Hard/Random score for this artifact is
missing.
Do not substitute the 88.18% Easy / 88.32% Hard result: that belongs to the
separate Data Scaling checkpoint trained with an additional 500 randomized
demonstrations per task.
Download and load
1hf download StarVLA/Qwen3-VL-OFT-Robotwin2 \
2 --local-dir playground/Pretrained_models/Qwen3-VL-OFT-Robotwin2
3
4export CKPT=playground/Pretrained_models/Qwen3-VL-OFT-Robotwin2/checkpoints/steps_40000_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
Keep the run directory intact so StarVLA can find its configuration and
statistics. Use the matching RoboTwin adapter and new_embodiment key.
Intended use and limitations
This release targets the RoboTwin 2.0 Base simulation embodiment. Results do
not establish performance for the randomized Data Scaling setting, alternative
camera or joint orders, or physical hardware. The packaged .pt file requires
trusted PyTorch deserialization and the matching StarVLA code.
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 terms and applicable dataset terms still
apply.
