PapersMulti-sourceSelected

ScaleVLA Scales Vision-Language-Action Models to 552B Parameters

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

Why we picked it

The non-obvious takeaway is that scaling the VLM backbone pays off most where small models struggle: contact-rich, precise manipulation and low-data regimes, not just high-level reasoning. Robotics labs and embodied-AI teams can use the controlled 2B-to-552B comparison to decide backbone size against latency budgets, since 216 ms per action chunk shows frontier-scale policies are deployable. The paper is on arXiv for anyone building on LIBERO-PRO benchmarks or real-world insertion tasks.

Researchers introduce ScaleVLA, a vision-language-action model that pairs a 552B DeepSeek-V4.1-Flash backbone with a trainable action diffusion transformer (DiT), and systematically compare VLM backbones from 2B to 552B under a shared training setup. The work tests whether frontier-scale VLM representations, already shown to help robot control in models like GPT-6 Astra, provide a stronger foundation for robot policies than the small backbones most VLA systems currently use.

Body

The evaluation covers three core capabilities of robot foundation models. In sample efficiency, ScaleVLA trained with just 10 demonstrations outperforms a 2B backbone trained with 50 (87.8% vs. 85.6%). In generalization, it reaches 45.1% success on LIBERO-PRO, versus at most 31.3% for smaller backbones. In precise manipulation, real-world charger insertion improves from 4/10 to 9/10 successes, with gains concentrated in contact-sensitive alignment. Despite its scale, the system achieves 216 ms median latency per action chunk.

The results suggest frontier VLMs can serve as representational foundations for efficient, generalizable robot policies rather than only reasoning agents, pointing toward larger backbones as a viable direction for VLA research.

Original · arXiv: DeepSeek

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

Frontier Vision-Language Models (VLMs) such as GPT-6 Astra have demonstrated remarkable capabilities in robot control, suggesting that broad pretraining provides knowledge valuable for physical action. Yet most Vision-Language-Action (VLA) policies still rely on much smaller VLM backbones, leaving this potential largely untapped. We ask whether frontier-scale representations can provide a stronger foundation for robot learning. We introduce ScaleVLA, which connects a 552B DeepSeek-V4.1-Flash backbone to a trainable action DiT, and systematically compare VLMs from 2B to 552B under a shared training setup. We evaluate three fundamental capabilities of robot foundation models. In sample efficiency, just 10 demonstrations outperform a 2B backbone trained with 50 (87.8% vs. 85.6%). In generalization, LIBERO-PRO success reaches 45.1%, compared with at most 31.3% for smaller backbones. In precise manipulation, real-world charger insertion improves from 4/10 to 9/10 successes, with gains concentrated in contact-sensitive alignment. Despite its scale, ScaleVLA achieves 216 ms median latency per action chunk. Our findings suggest that frontier VLMs can serve not only as reasoning agents, but also as powerful representational foundations for efficient and generalizable robot policies.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include ScaleVLA, DeepSeek-V4.1-Flash, VLA, robots.

Primary source verification: published directly by arXiv: DeepSeek with first-party authority. Cross-source corroboration: 3 independent sources have covered this development (including pandaily, 爱范儿), confirming industry-wide signal strength.