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
- HKU's TILDE Distills Embodied Foundation Models into Robot Policies — hkuhk · 2026-10-23
- 10 of 12 Beta Robots Worked Without Fine-Tuning — Do We Still Need It? — micoolcho · 2026-10-23
- EmbodiedEvo lets robot agents self-train, banking reusable Skills and Lessons across tasks — jiqizhixin · 2026-10-23
- KAIST's FRED lifts robot policy success from 19.4% to 50% via self-distilled future video representations — kaist-ai · 2026-10-23
- GALAX Launches World's First RTX 5080 With Dual 12V-2x6 Connectors at $3000 — chemist_slime · 2026-10-23
- Figure prints 81 minutes of robot data per second, spent ~$8.1M on data last week — joecole · 2026-10-23