Does VLA Action Prediction Need an LLM? TurboVLA Achieves 32 Hz with 0.2 B Params on RTX 4090

Vision‑Language‑Action (VLA) models are becoming a key technology for general robot control, but as these models grow larger and more semantically capable, their action‑update speed often cannot keep up. Existing VLA pipelines place a large language model (LLM) at the core: visual observations are first fed into a multi‑billion‑parameter LLM, fused with language instructions, and only then are robot actions generated. This design incurs substantial compute and memory overhead and prevents low‑latency closed‑loop execution.

The authors pose a central question: for robot tasks with explicit operation instructions, must each action prediction run a full LLM? To answer this, they introduce TurboVLA, a real‑time VLA model co‑developed by Huazhong University of Science and Technology and Huawei. TurboVLA removes the V→L→A pathway and instead lets visual and language features interact directly, forming a V+L→A pathway. The model uses compact encoders—DINOv3 for multi‑camera visual input and a lightweight BERT‑style encoder for language—followed by a bidirectional visual‑language interaction module that aligns task‑relevant visual regions with language cues. The resulting multimodal representation is combined with robot state and fed to a lightweight action decoder that outputs a full block of continuous actions in a single forward pass, eliminating the need for token‑wise generation or LLM‑based decoding.

TurboVLA’s efficiency is demonstrated on a single RTX 4090 GPU: the model occupies only 0.9 GB of VRAM, incurs a policy latency of 31.2 ms (≈32 Hz), and contains merely 0.2 B parameters. In the LIBERO benchmark (40 language‑conditioned tasks), TurboVLA achieves a 97.7 % average success rate, surpassing the 96.9 % of the larger π0.5 model (3.4 B parameters, 93.6 ms latency) while using roughly 6 % of the parameters and one‑third the latency.

On the RoboTwin 2.0 suite (50 clean dual‑arm tasks), TurboVLA attains a 60.2 % average success rate, outperforming π0.5 (57.0 %) and StarVLA‑α (50.3 %). The model size here is 0.4 B parameters with a 43.4 ms latency on RTX 4090. Real‑robot experiments on an AgileX Piper arm for tasks such as grabbing a roller, pressing a stapler, and stacking bowls yield success rates of 92.5 %, 80.0 %, 90.0 % and 87.5 % respectively, each exceeding the π0.5 baseline.

Ablation studies confirm the importance of language: removing language drops success to 70.8 %; replacing natural language with a task ID recovers performance to 95.4 % but remains below the full semantic instruction result. Moreover, simple concatenation of visual and language features reaches only 95.2 % success, whereas the bidirectional interaction module achieves the full 97.7 %, demonstrating that lightweight language representations combined with explicit cross‑modal interaction retain the essential semantic information for task execution.

In summary, TurboVLA shows that large language models are not mandatory for every action prediction step when the task instruction is already explicit. By allowing visual and language modalities to interact directly before action decoding, TurboVLA dramatically reduces model size, latency, and memory usage while preserving competitive robot manipulation performance. The authors suggest a hierarchical future where LLMs handle high‑level planning and TurboVLA handles low‑latency execution.