VLAct project pagePreprintAugust 2026

Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models

Core Contributors Senqiao Yang^†, Chengyao Wang^†, Yuxin Chen, Zixuan Wang, Longxiang Tang, Haokun Gui, Jinhui Ye, Changsheng Lu, Xiaoyang Wu, Mingkang Zhu † Project Leaders

Advisors Pengguang Chen, Shu Liu^✉, Zhuotao Tian, Hengshuang Zhao, Bei Yu, Jiaya Jia ✉ Correspondence

Make every robot trajectory teach more.

We hope to put a deeper question at the center of VLA research: what makes a VLM backbone a strong foundation for action in the physical world?

A Franka robot completing a long-horizon table-cleaning task — 01 Long-horizon execution

Two Franka arms coordinating during breakfast preparation — 02 Dual-arm coordination

Two Franka arms folding a pair of pants — 03 Deformable objects

Scroll to explore

Franka manipulation AgileX bimanual control GR-1 fine-tuned transfer ARX X5 cross-embodiment generalization Multiple action heads Open-data continued pre-training

01 · The premise

Robot data is scarce.
Representation is leverage.

“Under a fixed robot-data budget, VLA continued pre-training must convert limited trajectories into transferable visual-action knowledge, rather than merely fit actions.”

Large-scale robot trajectories cannot simply be scraped from the web. They are costly to collect, tied to particular bodies, and sparse relative to the physical world's diversity.

VLAct treats continued pre-training as representation learning—not simply action fitting—and is designed to preserve the broad VLM prior while acquiring transferable action semantics.

82.6%

LIBERO-Plus

Reported Total success rate, 7.6 percentage points above the matched Qwen3VL-OFT baseline.

92.5%

RoboTwin 2.0

Default VLAct-OFT Clean success in the Data Scaling setting, with 90.8% under randomization.

20%

RoboCasa-GR1 data

Of RoboCasa-GR1 fine-tuning trajectories—enough to exceed the reported 100%-data baselines shown below.

16 GPUs

Continued pre-training scale

VLAct continued pre-training uses open datasets in a 16-GPU setup.

02 · The result to remember

A robot embodiment unseen in continued pre-training—yet strong in transfer.

Robot continued pre-training uses Franka data from DROID and MolmoAct, plus AgileX data from InternData-A1 and RoboCoin. GR-1 trajectories are absent from continued pre-training; VLAct is then fine-tuned on RoboCasa-GR1. Move the data budget to inspect that adaptation.

20% of RoboCasa-GR1 trajectories

49.5% VLAct success

Fine-tuning data budget

Values are task success rates. Fractions refer to RoboCasa-GR1 fine-tuning trajectories; this evaluates transfer to an embodiment held out from continued pre-training, not zero-shot deployment.

RoboCasa-GR1 task success rate (%)

Higher is better

VLAct 20% fine-tuning data

49.5

reported 100%-data baselines

Qwen3VL-OFT100% fine-tuning data

48.8

GR00T-N1.6100% fine-tuning data

47.6

π0.5100% fine-tuning data

37.0

03 · Representation-centric continued pre-training

Before scaling more,
get the representation right.

Three continued-pretraining principles shape a more transferable backbone.

Preserve

Keep what the VLM already knows.

The entire vision encoder and lower half of the LLM layers are frozen during VLA continued pre-training, while caption supervision anchors broad visual and semantic knowledge. The full model is unfrozen again during downstream fine-tuning.

82.6%highest LIBERO-Plus result in Figure 8's auxiliary co-training ablation, achieved with caption data

Overview of the VLAct continued pre-training and fine-tuning recipe — Continued-pretraining heads act as scaffolding. Downstream tasks receive the backbone and a freshly initialized action head.

04 · Evidence across regimes

Transfer is the real test.

Controlled backbone comparisons keep the downstream head, data, optimizer, and fine-tuning budget fixed. Published systems are shown for broader context and are not controlled comparisons.

Robust single-arm manipulation

82.6%

Main-recipe result in the paper's LIBERO-Plus comparison

VLAct improves over the matched Qwen3VL-OFT baseline by 7.6 percentage points and surpasses ABot-M0 in the reported comparison.

Reported Total across seven perturbation axes; it is not a simple mean of rounded axis values

Reported Total success rate (%)

05 · Beyond simulation

From precise picks to
coordinated manipulation.

Physical setup: stationary, table-mounted Franka Research 3 arms; one fixed external Intel RealSense D435 for the setup and one wrist-mounted D405 on each arm. Each single-arm task uses 50 demonstrations, and each dual-arm task uses 100. Separate models are fine-tuned for single-arm and dual-arm evaluations, each for 50k steps on 8 H800 GPUs. Every task is evaluated over 10 rollouts.

Real robot · model rollout

Cube stacking

VLAct · single arm, short horizon

Choose a rollout

11 VLAct task videos · 2 matched baseline videos

Short horizon

Long horizon

Dual arm

Baselines

92.5%

Reliable short-horizon control

Mean success across four in-domain single-arm tasks, compared with 77.5% for Qwen3VL-4B-OFT without VLA continued pre-training; 10 rollouts per task.

VLAct gain+15.0 pp

06 · Models & checkpoints

Choose your
starting point.

The continued-pretraining backbone is the recommended base for new embodiments, datasets, and action heads. Downstream fine-tuned checkpoints are listed below.

Browse the VLAct collection Dataset Preparation

Continued-pretraining backbone

VLAct · Qwen3-VL-4B

The intended general-purpose VLAct representation before downstream task fine-tuning. Use it as the default adaptation base for a new robot, dataset, or action head.

StarVLA/VLAct_Qwen3_Pretrain Open repository

Fine-tuned checkpoints

Pick the target benchmark or dataset.

Artifact heads are named on each card and can differ from the paper's headline evaluation: RoboTwin's 92.5% result uses OFT, while the released card below uses GR00T; the listed VLA-Arena and LIBERO-Plus releases use PI, while Tables 1 and 4 report OFT-head comparisons. The RoboDojo repository reports a separate scaled local evaluation; the paper's 7.60% success and 10.66 score come from the official 50-episode-per-task leaderboard snapshot.

07 · Read and cite

Beyond data scaling.

Representation quality is an independent axis of VLA progress. Read the full report for protocols, ablations, per-task results, and implementation details.

Download PDF Project contact

Provisional BibTeX

@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/}
}

Replace with the official arXiv or venue entry when it becomes available.

Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models