ScaleVLA scales VLA models to 552B parameters: 10 demos beat 2B backbone trained on 50
GalaxyGeneralRobotics · hf · 2026-11-17
Teams from Tsinghua, Peking University, SJTU, and Galbot introduce ScaleVLA, connecting a 552B DeepSeek-V4.1-Flash backbone to a trainable action DiT, and systematically compare VLM backbones from 2B to 552B under a shared setup.
- Sample efficiency: just 10 demonstrations outperform a 2B backbone trained with 50 (87.8% vs 85.6%).
- Generalization: LIBERO-PRO success reaches 45.1% vs at most 31.3% for smaller backbones.
- Precise manipulation: real-world charger insertion improves from 4/10 to 9/10, gains concentrated in contact-sensitive alignment.
- Median latency is 216 ms per action chunk despite the scale.
Conclusion: frontier VLMs can serve as representational foundations for efficient, generalizable robot policies, not just reasoning agents.
More from Embodied
- XREAL Aura review: the spatial computer Apple should have made — Scobleizer · 2026-11-17
- Amazon's next delivery runner has four legs: RIVR robot joins the last-mile fleet — TinfoilTricorn · 2026-11-17
- Fudan and Fysics AI unveil UniWAM, hitting 80% success on real-robot mobile manipulation — jiqizhixin · 2026-11-17
- ChipAgents CEO joins NYSE interview on chip AI expansion and chip superintelligence vision — WilliamWangNLP · 2026-11-17
- Nvidia Robotics Chief Says OpenAI's Astra Makes Robotic Training Data 100x Faster — pstAsiatech · 2026-11-17
- Ant Group uses action-free egocentric human videos to boost robot policy learning — antgroup · 2026-11-17