TurboVLA: A Lightweight Vision-Language-Action Model Achieving 32 Hz Real-Time Inference on an RTX 4090

Forum topic · 小凯 · 2026-07-31

Summary

TurboVLA is a new vision-language-action (VLA) paradigm for robot control that replaces the conventional LLM-centric pathway — where visual observations are first projected into a large language model's representation space before decoding into actions — with a direct V + L -> A mapping. The model independently encodes visual observations and language instructions, exchanges information through lightweight bidirectional vision-language interaction, and predicts continuous action chunks with a compact decoder. This design builds task-conditioned representations directly from visual and language features, drastically reducing computation and memory costs per policy invocation. On the LIBERO benchmark, TurboVLA achieves a 97.7% average success rate with only 0.2B parameters, 31.2 ms inference latency, and 0.9 GB of GPU memory on a consumer-grade RTX 4090, matching or exceeding much larger VLA policies. The paper (arXiv 2607.27205) demonstrates that real-time, efficient VLA inference is achievable on consumer hardware.

论文概要

Research area: Computer Vision Authors: Hengyi Xie, Chenfei Yao, Xianjin Wu, Xuanyang Xi, Yiping Tang, Di Xu, Yingying Zhu, Dingkang Liang, Xiang Bai, Han Ding Published: 2026-07-29 arXiv: 2607.27205

Key points

Vision-language-action (VLA) models typically follow an LLM-centric V -> L -> A pathway: visual observations are projected into the representation space of a large language model, then decoded into robot actions. While effective, this design incurs substantial computation and memory overhead at every policy invocation.
TurboVLA reformulates the conventional V -> L -> A pathway as a direct V + L -> A mapping. Instead of using a large language model as the central interface between perception and action, it:
Independently encodes visual observations and language instructions
Exchanges information between them via lightweight bidirectional vision-language interaction
Predicts continuous action chunks with a compact decoder
This streamlined design builds task-conditioned representations directly from visual and language features, significantly reducing the computational and memory cost of VLA inference.

Results on LIBERO

| Metric | Value | |---|---| | Parameters | 0.2B | | Inference latency | 31.2 ms (~32 Hz) | | GPU memory | 0.9 GB | | Average success rate | 97.7% | | Hardware | Consumer-grade RTX 4090 |

TurboVLA matches or exceeds VLA policies with far more parameters, while running in real time on consumer hardware.