TurboVLA: A 0.2B-Parameter VLA Model Hits 32 Hz with <1 GB VRAM and 97.7% LIBERO Success on an RTX 4090

Forum topic · 小凯 · 2026-08-19

Summary

TurboVLA is a real-time vision-language-action (VLA) model that removes the large language model from the robot execution path. Using a compact 0.2B-parameter architecture with a lightweight language encoder and bidirectional cross-modal attention layers, it performs a direct vision+language-to-action mapping instead of the conventional vision-to-language-to-action pipeline. On LIBERO it achieves 97.7% average success across all four task subsets, with 31.2 ms inference latency (32 Hz) and 0.9 GB VRAM on a consumer RTX 4090—compared with OpenVLA's ~16 GB and 3–6 Hz. The 32 Hz control rate enables true closed-loop feedback control rather than open-loop replanning, and sub-1-GB memory permits 20+ parallel instances per GPU or edge deployment on Jetson-class hardware. The author argues TurboVLA defines a boundary: when instructions already specify the skill, open-ended language generation is redundant; LLMs remain necessary only for ambiguous, open-ended reasoning. Caveats include unverified sim2real transfer and real-world sensing latency not accounted for in the 31.2 ms figure. Paper: arXiv 2607.27205.

A 0.2B-parameter model runs at 31.2 ms inference latency with 0.9 GB VRAM on a consumer RTX 4090 and reaches 97.7% average success on LIBERO—while dismantling the assumption that a VLA model must have an LLM at its core.

Key numbers

| Metric | TurboVLA | Comparison | |---|---|---| | Parameters | 0.2B | OpenVLA 7B, RT-2 55B | | Inference VRAM | 0.9 GB | OpenVLA ~16 GB | | Inference latency | 31.2 ms (32 Hz) | OpenVLA ~150–300 ms (3–6 Hz) | | LIBERO average success | 97.7% | OpenVLA ~45%, RT-2 ~57%, π0 65–70% |

Paper info

Title: *TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM*
arXiv: https://arxiv.org/abs/2607.27205
HuggingFace: https://huggingface.co/H-EmbodVis/TurboVLA
Authors: H Xie et al., H-EmbodVis team
Published: July 2026

What TurboVLA does

The dominant VLA paradigm is V→L→A: a vision encoder translates images into language tokens, an LLM fuses them with the instruction to generate action tokens, and an action decoder outputs joint angles or end-effector poses. The problem: every inference wakes up a 7B+ LLM even when the instruction is an already-specified skill like "put the red block in the blue bowl."

TurboVLA instead uses a V+L→A direct mapping:

Vision encoder + lightweight language encoder (BERT-scale)
Bidirectional cross-modal interaction layers (vision and language tokens attend to each other)
Direct action output

Key claim: when the instruction already specifies the skill, the execution policy does not need open-ended language generation—the instruction only needs to decide how visual evidence guides action. The LLM's role on the execution path is redundant.

Three counterintuitive points

1. 97.7% on LIBERO is not a "small model, easy task" win. LIBERO's Spatial/Object/Goal/Long subsets test spatial understanding, object properties, goal orientation, and long-horizon tasks. OpenVLA-OFT and RDT-1B score only 23–29% on some subsets. TurboVLA's 97.7% average means near-perfect performance on all four.

2. 32 Hz is the hard threshold for real-time closed-loop control. Stable closed-loop feedback requires roughly 20–30 Hz. OpenVLA's 3–6 Hz only supports open-loop operation with periodic replanning. At 32 Hz, TurboVLA adjusts actions every 31 ms based on fresh visual feedback—a qualitative shift from "runs" to "runs in real time."

3. Sub-1-GB VRAM changes deployment economics. OpenVLA's ~16 GB means one GPU runs one instance. TurboVLA's 0.9 GB allows 20+ parallel instances on a single RTX 4090, or running VLA + visual SLAM + motion planning + safety monitoring together. It also enables edge deployment on platforms like Jetson Orin Nano (8 GB unified memory).

Observations

1. "De-LLM-ification" is a hidden theme in current AI system design. When task boundaries are clear and instructions are explicit, the LLM's open-ended generation is pure overhead. TurboVLA joins a trend of pulling LLMs back from "present everywhere" to "present only when needed."

2. Boundary conditions of the V+L→A mapping. The claim holds when instructions already specify skills. LIBERO's structured instructions flatter this. For ambiguous, open-ended tasks ("tidy up the desk," "something seems off, take a look"), the V→L→A paradigm with an LLM core may still be necessary. TurboVLA defines where direct mapping applies—explicit instructions, closed skill sets, short horizons. The LLM's battlefield shrinks; it does not disappear.

3. Bidirectional cross-modal interaction vs. one-way translation. V→L→A assumes vision must first be translated into language tokens—a lossy step. TurboVLA's bidirectional attention fuses vision and language directly in a shared multimodal representation, echoing Flamingo and BLIP-2 but pushed to the execution level.

4. Fitting the broader pattern. Like LocateAnything 3B showing perception does not need LLM-scale parameters, and neuro-symbolic work splitting "fast/fuzzy" from "correct/certain," TurboVLA shows vision-action mapping also does not need LLM scale.

A warning

LIBERO is a simulation benchmark. Sim2real gap is a chronic robotics problem—97.7% in sim might drop to 60–70% on real hardware, and the paper does not report detailed sim2real transfer. On a real robot, the perception pipeline (camera exposure + ISP + USB transfer + preprocessing) may consume 10–15 ms, leaving far less than 31.2 ms for the model. End-to-end latency and success on real hardware is the final verdict—don't judge by 97.7% and 32 Hz alone.

Open question

TurboVLA's bidirectional cross-modal layer is essentially a lightweight attention bridge. How compatible is it with LoRA fine-tuning? Could an ES-based (EGGROLL-style) fine-tune produce a TurboVLA-style lightweight action head on a pretrained OpenVLA model, reusing its visual representations? If so, this "VLA distillation" path might land faster than training TurboVLA from scratch.