Skip to content

Source paper

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

Ruixi Yu, Dairu Liu, Xuchuan Chen and 9 more

Ruixi Yu, Dairu Liu, Xuchuan Chen, Zekun Qi, Zekai Li, Zhuo Cao, Luyang Niu, Yingsheng Zhang, Sikai Liang, Ke Fan, He Wang, Li Yi

arXiv 2611.21607 · 202636 upvotes on Hugging Face

Research ideas from this paper (5)

Generated by AI. The paper’s authors did not write these.

Cross-Scale Representation Distillation from 552B MoE VLMs into Single-GPU Student Policies

Distill causal representations from frontier mixture-of-experts vision-language models into compact single-GPU student backbones so physical robots achieve high manipulation accuracy without multi-GPU hardware.

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters
🟠Medium💡Moderate InnovationLarge clusterRelated work found#Robotics#Efficiency

Parameter-Efficient Action Adaptation and Multi-Layer Feature Readout for Frontier VLMs

Adapt frozen frontier vision-language backbones via parameter-efficient fine-tuning and multi-layer feature aggregation to improve robot performance on out-of-distribution objects and contact-rich dexterous assembly.

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters
🔴Hard💡Moderate InnovationLarge clusterNo close match found#Robotics#Multimodal

Token-Normalized and Convergence-Controlled Evaluation of VLM Scaling Laws in Robotics

Isolate visual token count, convergence training budgets, and subset-only normalization to measure the true contribution of backbone parameter scale in robotic manipulation policies.

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters
🟡Fairly easy🔧Minor TweakLarge clusterNo close match found#Data & evaluation#Robotics

Dual-Rate Hierarchical Control with Closed-Loop Residual Correction for Contact-Rich Bimanual Manipulation

Combine asynchronous frontier vision-language feature updates with a high-frequency closed-loop residual controller to handle contact-sensitive alignment and observation staleness in bimanual manipulation.

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters
🔴Hard🔬Significant ResearchLarge clusterRelated work found#Robotics#Agents

Feature-Parity Verification and Latency-Accuracy Co-Evaluation for Served Frontier VLAs

Verify numerical feature parity between fast serving engines and evaluation pipelines while logging observation age during asynchronous robotic control.

ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters
🟡Fairly easy🔧Minor TweakMultiple GPUsRelated work found#Efficiency#Data & evaluation

AI summary

This paper investigates whether scaling the vision-language model (VLM) backbone of a vision-language-action (VLA) policy to frontier scale (552B parameters) improves robotic manipulation. By connecting a frozen 552B sparse mixture-of-experts VLM (DeepSeek-V4.1-Flash-1017) to a trainable flow-matching diffusion transformer action head, the authors demonstrate substantial gains in sample efficiency, out-of-distribution generalization, and contact-rich real-world insertion while maintaining subsecond per-chunk latency.

How it works (AI summary)

ScaleVLA pairs a frozen DeepSeek-V4.1-Flash-1017 vision-language backbone with a randomly initialized GR00T N1.7-style action head trained via flow matching to output continuous action chunks. Instead of autoregressive text reasoning, the model executes a single forward pass through the VLM's vision encoder and 20-layer causal encoder—stopping before the 20-layer decoder—and feeds the resulting 5120-dimensional visual and text token sequence into a feature adapter and a 32-layer alternating self- and cross-attention diffusion transformer over four integration steps.