TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on RTX 4090

H-EmbodVis · hf · 2026-07-30

Existing Vision-Language-Action (VLA) models typically adopt an LLM-centric pathway, projecting visual observations into the language representation space before decoding them into actions. This incurs substantial computation and memory overhead.

To address this, researchers introduced TurboVLA, a new paradigm that reformulates the conventional pathway into a direct "Vision + Language to Action" mapping. Instead of using an LLM as the central interface, TurboVLA independently encodes visual and linguistic inputs, exchanges information via lightweight bidirectional interaction, and predicts continuous action chunks with a compact decoder.

On the LIBERO benchmark, TurboVLA achieves a 97.7% average success rate using only 0.2B parameters. On a consumer-grade RTX 4090, it delivers a 31.2 ms inference latency (32 Hz) and requires less than 1 GB of VRAM, matching or outperforming significantly larger VLA policies.

Original post →

More from Embodied

Embodied channel →