TurboVLA replaces LLM with BERT, speeding robot actions

Researchers from Huazhong University of Science and Technology and Huawei introduced TurboVLA, a compact vision-language-action model that generates robot actions in 31.2 ms on an RTX 4090 and reaches 97.7 % success on the LIBERO benchmark. The trick is simple: the model contains no large language model (LLM). Instead, a BERT-based text encoder and a cross-attention module pass image and instruction information directly to the action head.

In classic VLA pipelines, visual features are projected into the token space of a language model, concatenated with instruction tokens, processed by a heavyweight LLM, and finally decoded into actions. The authors observed that when the instruction already specifies the desired manipulation (“stack three bowls”), the text only needs to guide visual attention. TurboVLA encodes image and instruction separately, exchanges them via cross-attention, and emits a chunk of 12 control steps in a single forward pass.

On LIBERO, TurboVLA achieves 97.7 % average accuracy with 0.2 B parameters, 0.9 GB VRAM and 31.2 ms latency. By comparison, π0.5 reaches 96.9 % with 3.4 B parameters, 12.8 GB VRAM and 93.6 ms; VLA-JEPA scores 97.2 % with 2.8 B parameters and 108.7 ms; CogVLA records 97.4 % with 8.3 B parameters and 115.5 ms. Among lightweight models, Evo-1 reports 94.8 % with 0.8 B parameters, VLA-Adapter 97.3 % with 1.5 B parameters. TurboVLA leads in accuracy while being the smallest and fastest of the group.

The team evaluated TurboVLA on an AgileX Piper robot across four tasks—lifting a cylinder, moving a card, pressing a button, and stacking three bowls. Success rates ranged from 80 % to 92.5 %, and TurboVLA outperformed π0.5 on every task. The simulation portion is fully reproducible: training and evaluation used open datasets LIBERO and RoboTwin 2.0 (the RoboTwin version employed a larger visual encoder of 0.4 B parameters). Proprietary data were used only for fine-tuning on the physical robot.

TurboVLA is designed for concrete instructions. It cannot decompose high-level commands such as “prepare breakfast” because it lacks an LLM capable of planning. The trade-off is explicit: speed and model size are gained at the expense of linguistic generalization.

The code is released on GitHub under the Apache-2.0 license, and the model weights are hosted on Hugging Face. The weights are “open” in the practical sense: they can be downloaded, run, fine-tuned and integrated into commercial products without additional permission. Full training data and the complete data-generation pipeline have not been published, so the release does not constitute full open-source reproducibility. For users needing a VLA model that runs on consumer-grade hardware without sacrificing accuracy, TurboVLA currently represents the most practical option.