VLAct: 16 GPUs, 20% Data Beats GR00T N1.6 in Cross-Embodiment Transfer

Beyond Data Scaling: Representation-Centric Continued Pre-training for VLA

Robotics VLA models have followed the LLM scaling paradigm—more data, larger models, more compute—but robot trajectories cannot be crawled from the internet like web data. Each trajectory requires physical robot operation, teleoperation, and hardware cost, making data expensive and scarce. VLAct addresses a complementary axis: improving representation quality so each trajectory yields more transferable knowledge.

The Core Problem: Head-Specific Representation Collapse

Experiments revealed a counter-intuitive phenomenon: improving action prediction on the pre-training head does not guarantee a more generalizable backbone. Using OFT head for continued pre-training raised downstream OFT success from 61.7% to 75.8% on RoboTwin, but switching the same backbone to PI head dropped performance from 60.5% to 55.1%, and to GR00T head from 51.2% to 28.9%. The backbone learns a feature geometry convenient for the specific head, a phenomenon termed head-specific representation collapse .

VLAct's Three Design Principles

Preserve VLM Priors : Freeze the vision encoder and shallow LLM layers; mix caption data during continued pre-training to retain original vision-language knowledge.

Multi-Head Continuous Supervision : Attach OFT, PI, and GR00T continuous action heads simultaneously, all reading the shared backbone and predicting the same actions. This forces the backbone to organize information in a head-agnostic way. After pre-training, all heads are discarded; downstream tasks re-initialize their own heads.

Partially Unified Action Space : Unify only truly consistent action dimensions across embodiments (e.g., gripper open/close semantics), while keeping embodiment-specific dimensions (arm joint spaces) separate.

Benchmark Results

RoboTwin 2.0 : 92.5% success (Clean), 90.8% (Random) under data scaling setup.

LIBERO-Plus : 82.6% success.

RoboDojo : 10.66 score, 7.60% success rate, surpassing all explicitly labeled World Action Models (WAM).

Real Robot Evaluation (Franka) : Covers single-arm short/long horizon, dual-arm coordination, novel objects, and full substitution OOD settings.

Cross-Embodiment Transfer: 20% Data Beats Full-Data Baselines

On RoboCasa-GR1 , the GR-1 robot never appeared in VLAct's pre-training data (which used Franka, AgileX, etc.). Using only 20% of RoboCasa-GR1 trajectories , VLAct achieves 49.5% success, already exceeding the full-data baselines of GR00T N1.6 and Qwen3VL-OFT. With 100% data, VLAct reaches 54.0%.

Training Efficiency and Open Release

VLAct continued pre-training starts from Qwen3-VL-4B , uses only public robot datasets and 16 GPUs . Model weights, checkpoints, data processing, and training pipeline are fully open-sourced.

Project: https://starvla.github.io/VLAct/ Paper: https://arxiv.org/abs/2608.27550 GitHub: https://github.com/starVLA/VLAct Hugging Face: https://huggingface.co/collections/StarVLA/vlact