News · 2026-07-30

A robot control model now runs 32 times a second on a gaming GPU, in under a gigabyte

A vision-language-action model called TurboVLA runs robot control at 32 decisions per second using 0.9 gigabytes of memory on a consumer RTX 4090 graphics card, with a 31.2-millisecond end-to-end policy latency. The model has 0.2 billion parameters. Its authors got there not by shrinking a large language model but by taking the language model out of the control loop altogether.

Key facts

• 0.2B parameters, 31.2 ms policy latency, 0.9 GB inference memory, 32 Hz control rate on a consumer RTX 4090.
• 97.7% average success on the LIBERO simulation benchmark; 80% to 92.5% on four real-world tasks on an AgileX Piper robot.
• Architecture: a DINOv3 vision backbone and a BERT instruction encoder feeding 6 bidirectional vision-language interaction layers, then a transformer decoder predicting continuous action chunks.
• Primary source: TurboVLA on arXiv, submitted July 29, 2026; code on GitHub.

The standard recipe for a robot policy today runs vision into a language model, has the language model reason, and derives actions from what it produces. That inherits everything good about large models - general knowledge, instruction following, the ability to handle a phrasing it has never seen - and one thing that is fatal for control: latency. A robot closing a loop on the physical world needs a decision every few tens of milliseconds. A large model thinking about each one cannot deliver that.

TurboVLA's argument is that the language model was never the part doing the controlling. The paper replaces the vision-to-language-to-action path with vision-and-language going directly to action: a vision transformer backbone and a compact text encoder feed six bidirectional interaction layers where the two modalities mix, and a decoder predicts a chunk of continuous actions in parallel. Training is plain behaviour cloning against demonstration data, with no auxiliary language-modelling objective at all.

The analogy is a skilled driver. When you take a corner you are not reasoning in sentences about the corner; the instruction "take the next left" entered your head as language once, and the actual control is a fast perceptual-motor loop that language is not in. TurboVLA keeps the instruction and deletes the internal monologue.

What that costs is stated clearly: the model gives up open-ended language generation and autonomous task decomposition inside the loop. It cannot be handed a vague goal and reason its way to a plan. It can be told what to do, in language, and do it very fast. For a system that pairs a slow planner with a fast controller - which is how most competent robot stacks are structured anyway - that is the right division.

The results are unusually well-covered on the axis that matters, which is whether it works on hardware. Simulation benchmarks are a low bar in this field, and TurboVLA clears LIBERO at 97.7%. More importantly it was fine-tuned on 65 teleoperated demonstrations per task and evaluated over 40 trials per task on a real AgileX Piper arm across four jobs - grabbing a roller, moving a playing card aside, pressing a stapler and stacking three bowls - with success rates of 92.5%, 80%, 90% and 87.5%. The loop closes on real hardware.

The number that will travel is the hardware requirement. Under a gigabyte on a graphics card that gamers already own puts real-time robot control inside the budget of a university lab, a startup, or a person with a workshop - the same shift that dropped the floor for running language models locally, arriving in robotics. It also lands the same week Google DeepMind announced Gemini Robotics 2 behind a waitlist, which makes for a sharp contrast: the frontier lab has the more capable system and a private preview; the open paper has weaker capability and a GitHub link.

The caveat is scope, and it is a large one. Four tabletop manipulation tasks on one arm is a narrow evaluation, and "press a stapler" is not "tidy this room." A companion paper released the same day, HumanCLAW, makes the limitation legible from the other direction: when nine frontier vision-language models are asked to find an object, walk their body to it and sit on it in simulation, the best manages under one in six episodes, and among the runs that reached the target, 71% still failed to sit - most of them by sitting into thin air. Fast control and competent embodied judgement are separate problems, and TurboVLA solves the first one.

Primary source, verified: read the paper → (arXiv 2607.27205)

Ground Truth. (2026, July 30). A robot control model now runs 32 times a second on a gaming GPU, in under a gigabyte. Ground Truth. https://groundtruth.day/news/a-robot-policy-that-runs-32-times-a-second-on-a-gaming-gpu.html

@misc{groundtruth:a-robot-policy-that-runs-32-times-a-second-on-a-gaming-gpu,
  title  = {A robot control model now runs 32 times a second on a gaming GPU, in under a gigabyte},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/a-robot-policy-that-runs-32-times-a-second-on-a-gaming-gpu.html}
}

Topics: robotics · vision-language-action · efficiency · local-inference · open-source · embodied-ai