TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

Researchers introduced TurboVLA, a compact vision-language-action model for robotic manipulation.

It skips the LLM-centered pipeline and uses direct visual-language interaction to predict actions with low latency.

TurboVLA reaches real-time robot control at 32 Hz on an RTX 4090 while using under 1 GB of VRAM.

The result points to a cheaper, faster way to run robot policies on consumer hardware.