Researchers from Hugging Face have unveiled TurboVLA, a novel vision-language-action (VLA) model that rethinks the conventional architecture. Instead of relying on a large language model (LLM) as the central bridge between perception and action, TurboVLA directly maps visual and linguistic inputs to continuous action chunks through lightweight bidirectional interactions.
The new design dramatically reduces computational and memory demands. On a consumer-grade RTX 4090 GPU, TurboVLA achieves an inference latency of just 31.2 milliseconds and uses less than 1 GB of VRAM, enabling real-time operation at 32 Hz. This is a significant improvement over existing VLA models that often require multiple GPUs or high-end hardware.
Benchmarked on the LIBERO robotic manipulation benchmark, TurboVLA attains a 97.7% average success rate with only 0.2 billion parameters, matching or surpassing much larger models. The authors emphasize that this efficiency does not come at the cost of performance, positioning TurboVLA as a practical alternative for real-world robotics applications.
The code and model are open-sourced on GitHub, allowing the research community to build upon this efficient VLA paradigm. The paper is available on Hugging Face Papers.