StarVLA-α: Reducing Complexity in Vision-Language-Action Systems

StarVLA-α is a deliberately minimal yet strong VLA baseline. It uses Qwen3-VL-4B-Instruct as its single vision-language backbone and pairs it with a small residual MLP that predicts a continuous action chunk in parallel. There is one model path, one supervised training loop, and one deployment interface.

🔴 StarVLA-α emerged from the systematic analysis enabled by the broader StarVLA platform. If you are interested in the complete engineering codebase, please visit the main StarVLA repository, where StarVLA-α has already been integrated.

✅ This repository is merged into StarVLA but remains open for issues and PRs. We will regularly merge accepted changes back to the main codebase.

StarVLA-α reduces VLA complexity while retaining strong cross-benchmark performance.

Why a minimal baseline?

Embodied AI sits at the intersection of robotics, vision, language, control, and data engineering. The many design choices and engineering tricks accumulated across these areas have made existing VLA codebases and baselines increasingly complex, creating a steep barrier for researchers and practitioners entering the field. StarVLA-α strips away that complexity and keeps only the essentials:

• The paper implementation standardizes on Qwen/Qwen3-VL-4B-Instruct, keeping the architecture and training recipe consistent across benchmarks.
• RGB observations and a language instruction are the model inputs; no state or history tokens are required.
• A shared MLP maps action-query hidden states directly to continuous actions.
• Training is a single-stage L1 regression objective on target robot data.
• Generalist training groups samples by native horizon and only pads the action dimension with a validity mask; it does not add another action architecture.
• The same registry, dataset discovery, config override, checkpoint, and policy-server contracts as the main StarVLA repository are retained.

The official StarVLA-α architecture, configs, released checkpoint, and reported results all use Qwen/Qwen3-VL-4B-Instruct. The loader retains Qwen2.5-VL and Qwen3.5 adapters for code-level compatibility with the main StarVLA repository, but substituting those backbones is an experimental extension rather than the StarVLA-α paper setting.

multi-view RGB + instruction
            │
            ▼
   Qwen3-VL-4B-Instruct
            │ action-query hidden states
            ▼
  shared residual MLP head
            │
            ▼
 continuous action chunk [T, D]

Paper-reported results

On the public RoboChallenge ARX5 evaluation (11 real-world tasks), the paper reports 33.6% success and 54.5 progress for StarVLA-α, versus 12.7% and 27.6 for π0.5. That is a 20.9 percentage-point success-rate improvement.

The released LIBERO 50k checkpoint has a separate model-card result: Spatial 99.0, Object 99.6, Goal 98.6, Long 94.8, average 98.0. Do not mix that checkpoint result with the paper's specialist table.

Installation

Python 3.10+ and a CUDA-capable PyTorch installation are required.

git clone https://github.com/starVLA/starVLA-alpha.git
cd starVLA-alpha
python -m pip install -e .

The standard installation above is the official Qwen3-VL setup. For code-compatibility experiments with the retained Qwen3.5 adapter, install the newer Transformers extra:

python -m pip install -e '.[qwen35]'

FlashAttention 2 is optional. If it is unavailable, the Qwen adapters fall back to PyTorch SDPA.

Released LIBERO checkpoint

Download the complete checkpoint directory and the 4B base model. The checkpoint loader intentionally requires config.yaml, dataset_statistics.json, and the weight file together.

hf download StarVLA/Qwen3-VL-OFT-LIBERO-4in1 \
  --local-dir playground/Checkpoints/Qwen3-VL-OFT-LIBERO-4in1

hf download Qwen/Qwen3-VL-4B-Instruct \
  --local-dir playground/Pretrained_models/Qwen3-VL-4B-Instruct

Start the policy server:

python deployment/model_server/server_policy.py \
  --ckpt_path playground/Checkpoints/Qwen3-VL-OFT-LIBERO-4in1/checkpoints/steps_50000_pytorch_model.pt \
  --config_override framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct \
  --use_bf16 \
  --port 6694

The server returns unnormalized action chunks and exposes the chunk length and normalization keys in its handshake metadata.

Checkpoint availability — The released LIBERO checkpoint above is available for evaluation. However, following a recent cluster migration, we have not yet recovered the exact checkpoints used to produce the results reported in the paper. We are working to secure available compute resources and reproduce them. We apologize for the inconvenience.

Training

The four paper benchmark configs differ only where the robot action space or data location requires it:

Benchmark	Config	Action shape
LIBERO	examples/simBenchmarks/LIBERO/train_files/starvla_qwenoft_libero.yaml	8 × 7
SimplerEnv	examples/simBenchmarks/SimplerEnv/train_files/starvla_qwenoft_simpler.yaml	16 × 7
RoboTwin	examples/simBenchmarks/Robotwin/train_files/starvla_qwenoft_robotwin.yaml	50 × 14
RoboCasa-GR1	examples/simBenchmarks/Robocasa_tabletop/train_files/starvla_qwenoft_robocasa.yaml	16 × 29
Generalist	scripts/train_files/starvla_qwenoft_generalist.yaml	native T ≤ 50, D padded to 32

Example, eight-GPU LIBERO training:

NUM_PROCESSES=8 bash scripts/train.sh examples/simBenchmarks/LIBERO/train_files/starvla_qwenoft_libero.yaml

All paper benchmark configs use Qwen/Qwen3-VL-4B-Instruct; changing the backbone produces an experimental variant outside the paper setting.

Generalist OFT

The generalist config composes the existing LIBERO, Bridge/RT-1, RoboTwin, and RoboCasa-GR1 registrations. There is no separate generalist framework or dataset fork. The YAML keeps one maximum horizon and one shared action width:

framework:
  action_model:
    action_dim: 32       # D padding bound
    action_horizon: 50   # T upper bound; native T remains dynamic

Samples are grouped by their native horizon before the VLM forward pass, so LIBERO, RoboCasa-GR1, and RoboTwin use 8, 16, and 50 action-query tokens respectively. The shared residual MLP is token-wise and therefore accepts every native T without changing its weights. Only D is zero-padded to 32, and a validity mask excludes that dimensional padding from the L1 loss. If any sample exceeds either YAML bound, training stops with an explicit shape error.

Place or symlink the registered dataset directories directly under playground/Datasets/StarVLA-Generalist/. The names must remain exactly those declared by the four benchmark registries; no generalist-specific conversion is required.

Launch generalist training with the same trainer:

NUM_PROCESSES=8 bash scripts/train.sh

At deployment, PolicyServerWrapper derives the native dimension and horizon from the selected unnorm_key. StarVLA-alpha generates the native number of action-query tokens and removes only dimensional padding before action un-normalization. StarVLA-Alpha is the primary registry name, while QwenOFT remains an alias for main-repository YAML files and released checkpoints. The server metadata exposes max_action_shape and per-key action_shapes so clients can schedule the returned chunk correctly.

Benchmark simulators have independent environment requirements. Keep them outside the StarVLA-α environment and communicate with the policy through the WebSocket interface under each benchmark's eval_files/ directory.

Repository scope

This repository is the paper implementation, not the full Lego-like StarVLA research platform. It intentionally excludes alternative action heads, world-model backbones, VLM co-training, unrelated robot integrations, and their dependencies. Use the main StarVLA repository when you need those extensions.

Citation

If StarVLA-α or the original StarVLA platform is useful for your research, we would appreciate a citation:

@article{ye2026starvlaalpha,
  title   = {StarVLA-$\alpha$: Reducing Complexity in Vision-Language-Action Systems},
  author  = {Ye, Jinhui and Gao, Ning and Yang, Senqiao and Zheng, Jinliang and Wang, Zixuan and Chen, Yuxin and Chen, Pengguang and Chen, Yilun and Liu, Shu and Jia, Jiaya},
  journal = {arXiv preprint arXiv:2604.11757},
  year    = {2026}
}

@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}
}

License

MIT. See LICENSE.