Qwen3-VL OFT baseline for RoboTwin 2.0

This repository provides a baseline policy for RoboTwin 2.0 trained using a minimal visual-to-action configuration.

The goal of this release is to provide a simple and reproducible baseline for RoboTwin, using:

• RGB observations only
• direct absolute joint prediction
• no robot state input
• no action transformer
• no pretraining

Despite its simplicity, the model achieves strong performance across RoboTwin 2.0 tasks.

Model and action contract

The policy directly maps RGB observations (head + two wrist cameras) to 14D absolute joint positions. The saved config uses a 50-step action horizon and does not enable robot-state input.

Training recipe

Training dataset

The model is trained using the official RoboTwin 2.0 dataset.

The training data consists of demonstrations from 50 manipulation tasks, combining clean demonstrations and randomized demonstrations.

Clean Demonstrations

• 50 tasks × 50 trajectories per task

Randomized Demonstrations

• 50 tasks × 500 trajectories per task

The randomized demonstrations include environment perturbations to improve robustness.

Evaluation protocol

The model is evaluated on RoboTwin 2.0 using 50 manipulation tasks.

Each task is evaluated under two evaluation settings:

• Easy setting
• Hard setting

For each task and setting:

• 100 evaluation trials are performed
• The success rate is reported

Final results are computed as the average success rate across tasks.

Benchmark results

Download and evaluation

huggingface-cli download StarVLA/Qwen3-VL-OFT-RoboTwin2-All \
  --local-dir Qwen3-VL-OFT-RoboTwin2-All

CKPT="$PWD/Qwen3-VL-OFT-RoboTwin2-All/checkpoints/steps_140000_pytorch_model.pt"
CUDA_VISIBLE_DEVICES=0 python deployment/model_server/server_policy.py \
  --ckpt_path "$CKPT" --port 5694 --use_bf16 \
  --config_override framework.qwenvl.base_vlm=Qwen/Qwen3-VL-4B-Instruct

The RoboTwin evaluator must use the saved absolute-action convention, normalization statistics, camera order, and 50-step prediction contract. See the StarVLA RoboTwin guide.

Evidence and limitations

• Architecture and training settings were checked against config.yaml; the released 140k step was checked against the Hub file tree.
• The benchmark table and 100-trial-per-task protocol are retained from the original Card. This repository does not include raw rollout logs, so the aggregate cannot be independently reconstructed from the Hub files alone.
• Only the 140k weight is available; the 150k configured endpoint is not released here.
• The policy is tied to the saved three-camera, absolute-joint and normalization contract. Different camera ordering, action mapping, chunk execution horizon, or normalization can materially change results.
• Results cover RoboTwin 2.0 simulation only. Real-robot performance and safety are not established.

Citation

If you use this model or benchmark results in your research, please cite RoboTwin and this repository.

@article{community2026starvla,
  title         = {StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing},
  author        = {Community, StarVLA},
  journal       = {arXiv preprint arXiv:2604.05014},
  year          = {2026},
  eprint        = {2604.05014},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}