HUST & Huawei's TurboVLA Bypasses LLM for 32Hz Real-Time Robot Control

机器之心 · wechat · 2026-08-03

Existing Vision-Language-Action (VLA) models typically use Large Language Models (LLMs) as an intermediary between vision and action, resulting in high computational overhead and latency. To address this, Huazhong University of Science and Technology (HUST) and Huawei proposed TurboVLA, a real-time VLA model.

Core Architectural Innovation

TurboVLA bypasses the LLM interface, adopting a more direct V+L→A path:

Uses DINOv3 and lightweight text encoders (e.g., BERT) to process vision and language separately.
Establishes task-scene correspondence via bidirectional vision-language interaction.
Feeds the interacted features directly into a lightweight action decoder to output continuous action chunks in a single forward pass.

Performance & Efficiency

Ultra-low Latency: Achieves 31.2ms end-to-end policy latency (32Hz) on a single RTX 4090 with only 0.9GB VRAM.
Minimal Parameters: With only 0.2B parameters, it achieves a 97.7% average success rate on the LIBERO benchmark, outperforming the 3.4B parameter π0.5 (96.9%) while reducing latency to a third.
Real-World Validation: Demonstrates excellent performance in both bi-manual simulation and real robotic arm tasks.

This research proves that for explicit robotic tasks, action prediction doesn't always require running a full LLM. A layered system where LLM handles high-level planning and TurboVLA handles low-latency execution could be a more efficient technical route.

Original post →

HUST & Huawei's TurboVLA Bypasses LLM for 32Hz Real-Time Robot Control

More from Embodied