You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

StarVLA Qwen RoboTwin inference bundle

This release contains the StarVLA code snapshot used for training and six RoboTwin policy checkpoints. Training logs, W&B files, optimizer states, other intermediate checkpoints, datasets, caches, local credentials, and local absolute paths are excluded.

Models

Directory	VLM backbone	Training	Action checkpoint
`checkpoints/qwen35-2b-robotwin`	`Qwen/Qwen3.5-2B`	150,000 optimizer steps	about 6.27 GB
`checkpoints/qwen35-4b-robotwin`	`Qwen/Qwen3.5-4B`	150,000 optimizer steps	about 11.48 GB
`checkpoints/qwen3vl-4b-robotwin`	`Qwen/Qwen3-VL-4B-Instruct`	150,000 optimizer steps	about 10.92 GB
`checkpoints/smolvlm2-2b-robotwin-step150000`	`HuggingFaceTB/SmolVLM2-2.2B-Instruct`	checkpoint at 150,000 optimizer steps	about 5.32 GB
`checkpoints/qwen35-9b-robotwin-step150000`	`Qwen/Qwen3.5-9B`	checkpoint at 150,000 optimizer steps	about 20.05 GB
`checkpoints/qwen35-9b-pelican-robotwin-step150000`	Pelican VLM-finetuned `Qwen3.5-9B`, then full VLA training	checkpoint at 150,000 optimizer steps	about 20.05 GB

Each checkpoint directory preserves the layout expected by StarVLA:

checkpoints/<model>/
├── config.yaml
├── config.full.yaml
├── dataset_statistics.json
└── final_model/
    └── pytorch_model.pt

The portable config.yaml files reference public upstream VLM repositories. The base VLM is downloaded on first use, and the complete trained VLA state dict is then loaded with strict key matching. The Pelican policy was initialized from the merged VLM checkpoint qwen35vl_lora_sft_60k_30k_15k_raw_v1_8gpu_merged/checkpoint-8500; its complete post-VLA-training state is contained in the uploaded policy checkpoint, so the public Qwen3.5-9B repository is sufficient to construct the architecture before strict state-dict loading.

Download on another server

python -m pip install -U "huggingface_hub[cli]"
huggingface-cli download kesnick/vlabench \
  --repo-type model \
  --local-dir starvla-qwen-robotwin
cd starvla-qwen-robotwin

For a private repository, run hf auth login first or export a read token as HF_TOKEN.

StarVLA environment

The tested environment uses Python 3.11, PyTorch 2.6.0+cu124, torchvision 0.21.0, Transformers 5.3.0, Accelerate 1.5.2, DeepSpeed 0.16.9, flash-linear-attention 0.3.2, causal-conv1d 1.5.0.post8, and Triton 3.2.0.

Create an environment compatible with the destination CUDA driver, then install the code and inference dependencies:

conda create -n starvla python=3.11 -y
conda activate starvla

# Install the matching CUDA build of torch/torchvision first.
pip install torch==2.6.0 torchvision==0.21.0 \
  --index-url https://download.pytorch.org/whl/cu124

pip install -r requirements-inference.txt
pip install -e ./code --no-deps

If the destination server cannot build the linear-attention CUDA packages, clone the already tested conda environment with conda-pack instead. The environment must retain Transformers 5.3.0; the repository's historical code/requirements.txt pins 4.57.0 and must not be used for these Qwen3.5 checkpoints.

RoboTwin environment

Install RoboTwin 2.0 from its official repository in a separate robotwin environment. Then install the lightweight evaluation client dependencies:

conda activate robotwin
pip install -r code/examples/simBenchmarks/Robotwin/eval_files/requirements.txt

Apply the policy_ckpt_path patch documented in:

code/examples/simBenchmarks/Robotwin/README.md

The patch modifies only the external RoboTwin checkout so it forwards the selected StarVLA checkpoint to the policy client.

Smoke evaluation

The StarVLA policy server and RoboTwin simulator run in separate environments. The unified launcher starts both sides and cleans them up automatically:

export ROBOTWIN_PATH=/absolute/path/to/RoboTwin
export STARVLA_PYTHON=/absolute/path/to/envs/starvla/bin/python
export ROBOTWIN_PYTHON=/absolute/path/to/envs/robotwin/bin/python
export CUDA_VISIBLE_DEVICES=0

cd code
bash examples/simBenchmarks/Robotwin/eval_files/start_eval.sh \
  -m demo_randomized \
  -n qwen35_2b_smoke \
  -c ../checkpoints/qwen35-2b-robotwin/final_model/pytorch_model.pt \
  adjust_bottle

Replace the checkpoint directory with any entry in the model table to evaluate another policy.

Multi-GPU full evaluation

export ROBOTWIN_PATH=/absolute/path/to/RoboTwin
export STARVLA_PYTHON=/absolute/path/to/envs/starvla/bin/python
export ROBOTWIN_PYTHON=/absolute/path/to/envs/robotwin/bin/python
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

cd code
bash examples/simBenchmarks/Robotwin/eval_files/start_eval.sh \
  -m demo_randomized \
  -n qwen35_4b_full_eval \
  -c ../checkpoints/qwen35-4b-robotwin/final_model/pytorch_model.pt \
  all

The formal observation contract is three RGB views in [head, left wrist, right wrist] order, resized to 224 x 224, with absolute 14-dimensional action chunks of horizon 50. dataset_statistics.json is used for action unnormalization.

Included and excluded files

Included:

StarVLA source and RoboTwin evaluation scripts from commit 312fac890ab75b7651d2bc4f8f8c8dbb5e055184, including the local Qwen3.5 and evaluation compatibility changes.
Final model weights only.
Portable runtime configuration and dataset normalization statistics.
Reproduction configuration and environment requirements.

Excluded:

Training and evaluation logs.
W&B local data and API keys.
Intermediate 10k-step checkpoints and optimizer/DeepSpeed state.
RoboTwin datasets and videos.
Base Qwen model weights already available from their upstream Hugging Face repositories.
Git history, caches, editor settings, and Python bytecode.

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Robotics