StarVLA open-sources VLAct: VLA backbone trained on 16 GPUs beats NVIDIA GR00T N1.6 with 20% of data

机器之心 · wechat · 2026-09-06

StarVLA released VLAct, a VLM backbone for vision-language-action models whose continued pre-training uses only open data and 16 GPUs. It hits 92.5% success on RoboTwin 2.0, tops all World Action Models on RoboDojo, and beats NVIDIA GR00T N1.6's full-data baseline on the unseen GR-1 robot using just 20% of downstream data. Weights, checkpoints and training pipeline are fully open-sourced.

The key insight: policy improving under one action head doesn't mean the backbone generalizes — swapping heads can tank performance, a phenomenon the paper calls head-specific representation collapse. VLAct counters by preserving VLM priors with caption data, supervising with three action heads (OFT, PI, GR00T) simultaneously on one backbone (all discarded downstream), and unifying only physically consistent action dimensions across embodiments. Built on Qwen3-VL-4B, it argues representation quality deserves to be a scaling axis independent of data volume.

Original post →

More from Embodied

Embodied channel →