TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM
Hengyi Xie, Chenfei Yao, Xianjin Wu, Xuanyang Xi, Yiping Tang, Di Xu, Yingying Zhu, Dingkang Liang, Xiang Bai, Han Ding
cs.CV, cs.RO
2026-07-30
TurboVLA rewires VLA execution into a V+L to A path using DINOv3 and lightweight BERT; it runs at 32 Hz with 0.9 GB VRAM and 0.2B parameters on an RTX 4090, reaching 97.7% on LIBERO and beating π0.5.
Vision-language-action (VLA) models that drive robot arms share two flaws: slow and memory-hungry. The mainstream follows a V to L to A pathway that funnels images through a large language model before decoding actions, either autoregressively token by token like OpenVLA and RT-2 (slow) or skipping token generation but still routing everything through a multi-billion-parameter LLM core like π0 and π0.5 (heavy). On an RTX 4090 these range from 84 ms (π0) to 203 ms (SmolVLA), with VRAM often in the tens of gigabytes. A robot arm needs a control loop that returns an action within tens of milliseconds, and this architecture is structurally wrong for it. The paper's bet is direct: execution-level instructions do not need open-ended language generation, so the LLM middle layer can go.
TurboVLA rewires the pathway to V+L to A and drops the LLM entirely. Vision uses DINOv3 (a discriminative vision encoder strong on structured features) and text a lightweight BERT, with bidirectional cross-attention (borrowed from Grounding DINO) for task conditioning instead of a large-model bridge. The key judgment here: real-robot instructions are mostly deterministic phrases like "put the red block in the bowl," which do not need an LLM's generation or reasoning; BERT's semantic grasp is enough, and what gets cut is the inference cost of billions of parameters. Actions come from an ACT-style non-autoregressive transformer decoder that emits a continuous action chunk (12 steps on LIBERO, 50 for bimanual) in a single forward pass, with no tokenization and no sequential generation, which is the main reason latency drops to milliseconds. Robot state is encoded separately and injected only at the decoder, keeping the vision-text cross-modal interaction focused on scene understanding. The whole model is 0.2B parameters.
On an RTX 4090 at batch size 1, LIBERO (single-arm) reaches 97.7% average success at 31.2 ms (32 Hz), 0.9 GB VRAM, and 0.2B parameters. Against π0.5 (96.9% success, 93.6 ms, 3.4B parameters, 12.8 GB VRAM), TurboVLA uses 6% of the parameters and scores higher. Across the latency field, π0 sits at 84.2 ms, OpenVLA 202.9 ms, SmolVLA 203.1 ms, and VLA-JEPA 108.7 ms; TurboVLA is an order of magnitude below them all, about 3x faster than VLA-JEPA with 7% of its parameters. Bimanual (RoboTwin 2.0) hits 60.2% success at 43.4 ms with 0.4B parameters; real-world (AgileX Piper) scores 92.5%, 80%, 90%, 87.5% on four tasks.
For anyone deploying VLAs on real robots, this shows the big LLM in the execution loop is not required: remove it and latency drops to milliseconds, VRAM fits a consumer card, and parameter count shrinks by two orders of magnitude with no loss in task success. That buys higher-frequency control loops and cheaper hardware. The cost is in the next section.
TurboVLA handles concrete execution-level instructions well but cannot do the complex semantic understanding and high-level planning. The authors say planning is better left to an LLM and frame their path as efficient execution, with a combined LLM-planning-plus-TurboVLA-execution setup as the likely future. Real-world task variety is also still limited, and large-scale generalization is not validated here.