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.
More from Embodied
- Xiaomi humanoid robot logs 98% success on factory tasks with 66 DoF, half in the hands — Olivier__OG · 2026-09-06
- Robots get their GPT-3 moment: In-Context Learning lets them learn new tasks from one demo — 量子位 · 2026-09-06
- ETH Zurich open-sources full 2026 robot learning course: 12 weeks, VLA models, free — Syntetisaattori · 2026-09-06
- Yacine Builds His Own Eye-Tracking Hardware with Custom Drivers for High-FPS Tracking — yacineMTB · 2026-09-06
- Show Lab sweeps 3/3 at CoRL 2026: MetaWAM hits 68.9% on RoboCasa with 39% lower inference latency — MikeShou1 · 2026-09-06
- NSF Launches $30M Center for Human and Robot Co-Adaptation Led by UT Austin — PeterStone_TX · 2026-09-06