Megadose The AI curve, compressed.

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters

· HF Daily Papers ·
With 10 demonstrations, ScaleVLA achieved 87.8% success, beating a 2B model trained on 50 at 85.6%.

The system connects a 552B DeepSeek-V4.1-Flash backbone to a trainable action transformer, comparing backbones from 2B to 552B under a shared training setup. On LIBERO-PRO, it reached 45.1% success versus at most 31.3% for smaller backbones. Real-world charger insertion improved from 4/10 to 9/10 successes, with gains concentrated in contact-sensitive alignment. Median latency was 216 milliseconds per action chunk.

HF Daily Papers' note

score 5

Categories: Research