TurboVLA hits 97.7% on robot tasks, skips the LLM

TurboVLA scored 97.7% on the LIBERO robot benchmark without a large language model in the loop, per its arXiv paper.

It does this at 32 Hz on a single RTX 4090, using 0.9 GB of VRAM. That matters for anyone trying to run a robot control policy in real time without a data center behind it. Most vision-language-action systems route camera input and instructions through an LLM first, which adds compute. That slows response time.

TurboVLA skips that step. It maps vision and language straight to action using separate encoders and what the paper calls lightweight interaction between them. That replaces the usual vision-to-language-to-action chain.

TurboVLA uses just 0.2 billion parameters and runs with 31.2 millisecond latency. The paper's code is public, so the 97.7% score is one anyone can try to reproduce.

The paper doesn't say how TurboVLA performs on tasks or benchmarks outside LIBERO, or on physical hardware beyond the 4090 latency figure it reports.

If that 97.7% score holds up, it's an argument that the field's push toward bigger LLM backbones for robot control solved the wrong problem. A 0.2 billion parameter model matching benchmark scores that bigger vision-language-action systems chase suggests those extra billions were buying generality LIBERO's tasks don't need.

Worth watching whether TurboVLA gets tested on real robot hardware or on benchmarks tougher than LIBERO.

It topped HuggingFace's Daily Papers rankings on July 30, at 109 upvotes.