# TurboVLA’s 32 Hz metric is a 12-step action chunk—not a control loop Canonical URL: https://www.theroboteconomy.news/articles/turbovla-action-chunk-closed-loop-gap/ Publisher: The Robot Economy Editorial byline: Imani Reed, Frontier Desk Published: 2026-07-30T09:25:31.149Z Updated: 2026-07-30T09:25:31.149Z Editorial information: https://www.theroboteconomy.news/about/#editorial-desks The newly released code clarifies that the reported 31.2-millisecond VLA result produces 12 open-loop actions; four hardware tasks were fine-tuned on 65 demonstrations each. ## Summary TurboVLA is a newly released 0.2-billion-parameter vision-language-action model from Huazhong University of Science and Technology and Huawei. Authors report 97.7 percent average LIBERO success and 31.2 millisecond action-chunk inference. The code configures 12 open-loop steps per query. Four hardware tasks used 65 teleoperated demonstrations and 40 trials each. Training and evaluation code is public under Apache-2.0; weights and datasets are external assets. ## Reader context Vision-language-action model: A model that turns images and a language instruction into physical robot actions. Action chunk: A sequence of robot actions predicted together before the policy makes its next high-level decision. ## Key facts - TurboVLA reports 31.2 millisecond action-chunk inference on an RTX 4090 for its LIBERO result. - The released LIBERO recipe configures a 12-step chunk and 12 open-loop steps per action query. - The paper's four hardware tasks were each fine-tuned on 65 teleoperated demonstrations and evaluated over 40 trials. - The public repository released training and evaluation code under Apache-2.0, while model weights and datasets remain external assets. ## Lead image Image: https://www.theroboteconomy.news/images/articles/turbovla-action-chunk-closed-loop-gap.png Description: A generic tabletop robot arm beside a compact compute module, with blue sensor dots and a short amber segmented motion path. Credit: AI-generated editorial illustration / The Robot Economy ## Report ### What the 32 Hz result measures TurboVLA is a 0.2-billion-parameter vision-language-action model from Huazhong University of Science and Technology and Huawei that the authors say reaches 97.7 percent average success on LIBERO with 31.2 milliseconds of inference time and 0.9 gigabytes of inference memory on an RTX 4090. The paper reports 2,000 simulation trials across LIBERO’s four task suites. Its public repository released training and evaluation code on July 30 under Apache-2.0. Sources: [s1]() [s2]() [s3]() The decision delta is in the unit of speed. TurboVLA’s 31.2-millisecond number is the time to produce one action chunk, not a claim that the robot re-observes and replans 32 times a second. The released LIBERO recipe sets both the chunk size and open-loop execution length to 12 steps. That is still an important resource result for on-robot compute, but it changes the operational question from raw inference latency to how the policy behaves between high-level visual updates. Sources: [s1]() [s2]() [s3]() ### What the hardware test can establish The authors also tested an AgileX Piper arm on four tasks: grabbing a roller, moving a playing card, pressing a stapler and stacking bowls. Each task was fine-tuned with 65 teleoperated demonstrations and evaluated over 40 trials against pi-zero-point-five on the same platform. Those experiments support a bounded hardware result, not a general claim of out-of-distribution robustness, unattended operation or broad robot transfer. The real-world data and procedure are not part of the published code recipe. Sources: [s1]() [s2]() ### The reproduction checkpoint The code release makes the simulation path more inspectable: it names separate LIBERO and RoboTwin environments, data formats and evaluation commands. But the repository says model weights and benchmark datasets are external assets, and its documented recipes cover those simulation stacks rather than a downloadable TurboVLA checkpoint or hardware protocol. The next meaningful test is an independent team running the released approach on a physical task with raw sensory disturbance, longer horizons and disclosed intervention rules. Until then, the evidence supports an efficient benchmark-and-lab result, not deployment readiness. Sources: [s2]() [s3]() [s4]() ## Conclusion The article maps TurboVLA's appealing 32 hertz headline to the implementation that produces 12 open-loop actions at once. That source-level distinction separates a real inference-efficiency gain from the much stronger claim of continuous closed-loop robot reaction, while also locating the remaining reproduction gap in its released artifacts. ## Disclosures Sponsored: no Desk bylines, AI assistance and editorial policies: https://www.theroboteconomy.news/about/ ## Sources - s1: [TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with \<1 GB VRAM]() — arXiv; paper; 2026-07-29T17:59:58.000Z - s2: [TurboVLA official implementation]() — H-EmbodVis; code-release; 2026-07-30T00:00:00.000Z - s3: [LIBERO: Benchmarking Knowledge Transfer in Lifelong Robot Learning]() — Lifelong Robot Learning; code-release; 2023-09-22T00:00:00.000Z - s4: [RoboTwin 2.0 official repository]() — RoboTwin Platform; code-release; 2025-08-28T00:00:00.000Z