TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

TL;DR

TurboVLA (Huawei / HUST) reformulates the standard LLM-centric VLA pathway V → L → A (visual features projected into the LLM latent, decoded to actions) as a direct V + L → A mapping — the vision encoder and language encoder are independent, and information is exchanged between them via a lightweight bidirectional vision-language interaction module before a compact decoder emits continuous action chunks. The resulting policy is 0.2B parameters, uses ~0.9 GB inference VRAM, and runs at 31.2 ms per policy invocation (32 Hz) on a consumer RTX 4090. On LIBERO it reports 97.7% average success — matching or beating substantially larger VLA policies while eliminating the LLM as the central perception-to-action bottleneck.

Key claims

Reformulating the VLA pathway from V → L → A (visual observations routed through an LLM’s representation space before action decoding) to a direct V + L → A mapping removes the LLM’s per-invocation compute + memory overhead as a policy bottleneck [Abstract, §Introduction].
The load-bearing design piece is a lightweight bidirectional vision-language interaction module that independently-encoded visual and language streams use to exchange information, producing task-conditioned representations without an LLM central interface [Abstract].
A compact decoder predicts continuous action chunks directly from the task-conditioned representation, avoiding the LLM-decoder path used by most open VLAs [Abstract].
At 0.2B parameters TurboVLA reports 97.7% average success on LIBERO, matching or outperforming substantially larger VLA baselines [Abstract].
Reported deployment envelope on a consumer RTX 4090: 31.2 ms inference latency (≈32 Hz policy rate) and 0.9 GB inference VRAM [Abstract].
Code is released (github.com/H-EmbodVis/TurboVLA, Apache 2.0) [Abstract].

Method

TurboVLA discards the standard VLA architectural template of “run everything through a language-model backbone.” A vision encoder independently produces visual features from robot camera observations, and a separate language encoder independently processes the natural-language instruction. Instead of projecting vision into the LLM’s representation space and decoding actions autoregressively from a shared LLM, TurboVLA introduces a lightweight bidirectional vision-language interaction layer between the two encoders — visual features attend to language features and vice-versa in a symmetric fashion — so a task-conditioned representation is built up directly from both modalities without an LLM acting as central interface. A compact action decoder then reads this joint representation and emits continuous action chunks. The absence of an LLM central pathway is what collapses the parameter count to 0.2B and the memory footprint to <1 GB.

Results

LIBERO overall: 97.7% average success at 0.2B parameters — reported to match or outperform substantially larger LLM-centric VLA policies [Abstract].
Inference latency: 31.2 ms per policy invocation on a consumer RTX 4090, i.e. ≈32 Hz control rate [Abstract].
Inference memory: 0.9 GB VRAM — reported as enabling compact local deployment [Abstract].
The abstract does not enumerate per-baseline numbers or LIBERO subset splits; those would live in the tables of the full paper.

Why it’s interesting

TurboVLA lands on the same axis that Patch Policy: Efficient Embodied Control via Dense Visual Representations recently staked out — skip the LLM/VLM as central perception-to-action interface — but takes a different route: Patch Policy attaches a small transformer directly to a frozen SOTA ViT’s dense patch tokens, while TurboVLA keeps a language encoder in the pipeline and puts the emphasis on a bidirectional V↔L interaction module in place of an LLM latent bus. Both converge on the same downstream evidence — a sub-billion-parameter policy competitive with or beating multi-billion-parameter VLAs, at ~10-30 ms latency on a consumer GPU — which sharpens the argument on the VLA Models concept page that the LLM-centric pathway is a design choice, not a scaling law. TurboVLA is also a concrete counter to LingBot-VA-style unified LLM+action approaches (Causal World Modeling for Robot Control (LingBot-VA)) and to the action-pretraining-at-scale position of π*0.6: a VLA That Learns From Experience (RECAP): same LIBERO tier, ~10× smaller, no LLM. For the edge-deployment thread (Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots, Introducing Cosmos 3 Edge), a 0.9 GB / 32 Hz VLA is well below the memory and latency thresholds those runtimes were designed around.