StarVLA/VLAct_Qwen3PI_VLA_Arena_Finetune

VLAct · Qwen3-VL-4B PI · VLA-Arena

This repository contains the 50K-step VLAct downstream fine-tuning checkpoint for VLA-Arena, introduced in Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models. It starts from StarVLA/VLAct_Qwen3_Pretrain and adapts the shared VLAct backbone with a randomly initialized PI (QwenPI_v4) action head for the target benchmark.

This is a StarVLA training / evaluation checkpoint, not a standard transformers.AutoModel package. Load it with the matching StarVLA framework (QwenPI_v4) and the packaged config.yaml / dataset_statistics.json. Safe robot deployment still requires embodiment-specific action mapping, normalization, camera calibration, control-rate handling, workspace constraints, and independent safety systems.

What is VLAct?

VLAct is a representation-centric continued-pretraining recipe for vision-language-action models. It preserves the VLM prior, co-trains multiple continuous action heads on a shared latent, and shares action semantics across embodiments with a partially unified padded layout and wrap-aware joint loss. Downstream policies discard the pretraining heads, randomly initialize a target action head, and fine-tune from the VLAct backbone.

The complete method, ablations, and evaluation protocols are documented in the paper and code repository.

Checkpoint details

Item	Value
Framework	StarVLA QwenPI_v4
Base VLM	StarVLA/Qwen3-VL-4B-Instruct-Action
Pretrained backbone	StarVLA/VLAct_Qwen3_Pretrain @ 100K
Action head	PI / flow-matching, randomly initialized at fine-tuning start
Action dimensions	7-D
Action horizon	8 steps (future_action_window_size: 7)
Diffusion settings	num_inference_timesteps: 4, repeated_diffusion_steps: 2
Dataset mix	vla_arena_L0_L (Large split)
Data root	playground/Datasets/VLA-Arena
Training step	50,000
Per-GPU VLA batch size	16
Learning rates	Qwen-VL interface 1e-5; action/base modules 1e-4
Scheduler	cosine with min_lr: 1e-6, 4K warmup
Seed	42

Fine-tuning data

Fine-tuning uses the VLA-Arena L0 Large split (vla_arena_L0_L) with a CoT object-grounding prompt. VLA-Arena covers multiple evaluation domains and difficulty levels; see examples/VLA-Arena/README.md for the data layout and metrics.

Recommended use: download and evaluate

1. Install StarVLA / VLAct

git clone https://github.com/starVLA/VLAct.git
cd VLAct

conda create -n vlact python=3.10 -y
conda activate vlact

# Install a CUDA-compatible PyTorch build first.
python -m pip install -r requirements.txt
python -m pip install flash-attn==2.7.4.post1 --no-build-isolation
python -m pip install -e .

2. Download the checkpoint

Run from the VLAct repository root:

hf download StarVLA/VLAct_Qwen3PI_VLA_Arena_Finetune \
  --local-dir playground/Pretrained_models/VLAct-Qwen3VL4B-PI-VLA-Arena

The checkpoint path is then:

playground/Pretrained_models/VLAct-Qwen3VL4B-PI-VLA-Arena/checkpoints/steps_50000_pytorch_model.pt

Keep the downloaded directory structure unchanged. StarVLA resolves config.yaml and dataset_statistics.json from the run directory two levels above the checkpoint file.

Action un-normalization

The policy predicts actions normalized to roughly [-1, 1]. StarVLA maps them back to physical units using the q01 / q99 / mask statistics stored in dataset_statistics.json, and the unnorm_key selects which statistics block to use.

This run packages a single key, franka. StarVLA resolves it automatically when only one key is present, so you normally do not need to set anything. If your evaluation config exposes an unnorm_key field (for example examples/LIBERO-plus/eval_files/eval_libero.sh or the VLA-Arena client setup), set it to franka.

Changing the normalization statistics, camera ordering, state usage, action ordering, or execution horizon can materially change results.

3. Evaluate with the StarVLA / VLAct scripts

Follow the VLA-Arena guide:

examples/VLA-Arena/README.md

Reproduce training with:

bash scripts/run_scripts/VLA-Arena/train_vla_arena_qwen3pi.sh

Loading the policy

Reconstruct the policy with the matching StarVLA framework (QwenPI_v4) and the packaged configuration. The .pt file contains model parameters only; it does not package optimizer or scheduler state.

This checkpoint is intended for evaluation or further fine-tuning on the same embodiment / action contract. Transferring it to a different robot, camera setup, or action space usually requires additional adaptation.

Files

VLAct-Qwen3VL4B-PI-VLA-Arena/
├── README.md
├── config.yaml
├── training_config.original.yaml
├── dataset_statistics.json
├── summary.jsonl
└── checkpoints/
    └── steps_50000_pytorch_model.pt

File	Purpose
checkpoints/steps_50000_pytorch_model.pt	Fine-tuned PyTorch state dict for the downstream policy
config.yaml	Portable resolved configuration using the public base-model ID
training_config.original.yaml	Original resolved run configuration as produced by training; it records the internal base-model ID used at training time and is kept for provenance only
dataset_statistics.json	Dataset statistics used by StarVLA normalization utilities
summary.jsonl	Saved-checkpoint step history

Checkpoint SHA-256:

2e8c86aaa5295983e80a59b4f465d4eb6f0b2ca8c8c08a4a639f7754614892a9

Benchmark results

VLAct's published VLA-Arena result:

Setting	VLAct	Matched Qwen3-VL-OFT baseline
VLA-Arena	54.8%	33.4%

The released artifact head can differ from the head used in the paper's headline table for the same benchmark. The numbers above are the published VLAct results for this benchmark; they are not a re-evaluation of this specific checkpoint.

The paper's VLA-Arena comparison reports an OFT-head policy, while this released checkpoint uses the PI head. Treat the table as benchmark context for the VLAct recipe.

Metrics are averaged over 50 trials per task level (5 tasks x 10 episodes). See examples/VLA-Arena/README.md for the evaluation setup.

Intended use and limitations

This checkpoint is intended for research on VLA representation transfer and benchmark evaluation on VLA-Arena.

• It has not been validated as a universal zero-shot policy across arbitrary robots.
• Safe deployment requires embodiment-specific action mapping, normalization, camera calibration, control-rate handling, workspace constraints, and independent safety systems.
• Performance depends on evaluation protocol, simulator / real-robot setup, observation configuration, and action execution settings.
• The model may inherit limitations and biases from its base VLM, the VLAct pretraining mixture, and the downstream fine-tuning data.

Citation

@misc{yang2026vlact,
  title   = {Beyond Data Scaling: Representation-Centric Continued Pre-training
             for Vision-Language-Action Models},
  author  = {Yang, Senqiao and Wang, Chengyao and Chen, Yuxin and Wang, Zixuan and
             Tang, Longxiang and Gui, Haokun and Ye, Jinhui and Lu, Changsheng and
             Wu, Xiaoyang and Zhu, Mingkang and Chen, Pengguang and Liu, Shu and
             Tian, Zhuotao and Zhao, Hengshuang and Yu, Bei and Jia, Jiaya},
  year    = {2026},
  month   = aug,
  note    = {Preprint},
  url     = {https://starvla.github.io/VLAct/}
}

License and acknowledgements

The checkpoint is released under the Apache License 2.0. The VLAct code repository is released separately under the MIT License. Users must also comply with the licenses and terms of the base model, the VLAct pretraining checkpoint, and the training / evaluation datasets.

VLAct builds on StarVLA, LeRobot, GR00T, and Qwen3-VL.

For questions, email yangsenqiao.ai@gmail.com or open an issue in the VLAct repository.