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

Source-linked AI summary

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

arXiv:2607.27205v1Jul 29, 2026cs.CVcs.RO

TL;DR

LLM-centric VLA models impose substantial computation and memory costs that hinder real-time robotic execution. TurboVLA directly couples lightweight visual and language encoders with cross-modal interaction and action-chunk decoding, achieving 97.7% average LIBERO success at 31.2 ms latency with 0.2B parameters and 0.9 GB VRAM on an RTX 4090.

Problem
LLM-centric VLA execution pathways create substantial computation and memory bottlenecks for real-time robotic control.
Method
TurboVLA separately encodes vision and language, fuses them through lightweight direct interaction, and decodes continuous action chunks without an LLM-centered execution pathway.
Results
97.7% average LIBERO success accompanies 31.2 ms latency, 0.2B parameters, and less than 1 GB inference VRAM, matching or outperforming substantially larger VLA systems.
Takeaways & Limitations
TurboVLA validates direct V + L → A execution as a way to combine strong manipulation performance with low latency and compact model scale.

Abstract

from arXiv · show

Vision-language-action (VLA) models commonly adopt an LLM-centric $V \to L \to A$ pathway, where visual observations are projected into the representation space of a large language model before being decoded into robot actions. Although effective, this design incurs substantial computation and memory overhead at every policy invocation. In this work, we introduce TurboVLA, a new VLA paradigm that reformulates the conventional $V \to L \to A$ pathway as a direct $V + L \to A$ mapping. Instead of using a large language model as the central interface between perception and action, TurboVLA independently encodes visual observations and language instructions, directly exchanges information between them through lightweight bidirectional vision-language interaction, and predicts continuous action chunks with a compact decoder. This simple design constructs task-conditioned representations directly from visual and linguistic features, significantly reducing the computational and memory costs of VLA inference. On LIBERO, TurboVLA achieves 97.7% average success with only 0.2B parameters, 31.2 ms inference latency, and 0.9 GB inference VRAM on a consumer-grade RTX 4090, matching or outperforming substantially larger VLA policies. These results establish TurboVLA as a simple and effective alternative to the prevailing LLM-centric VLA paradigm, offering a new perspective on how vision, language, and action can be connected for efficient robotic manipulation. Code is available at https://github.com/H-EmbodVis/TurboVLA.

1 INTRODUCTION

TurboVLA replaces the conventional LLM-centered VLA execution pathway with direct vision-language interaction and lightweight instruction encoding, targeting real-time language-conditioned manipulation. On an RTX 4090, it achieves 32 Hz execution with 0.2B parameters and 31.2 ms latency while preserving strong manipulation performance.

LLM-centric VLA models create a substantial bottleneck for real-time robotic execution, especially in responsive interaction, high-throughput manipulation, and resource-constrained deployment.
Execution-level control need not use a large language model when instructions already specify the intended manipulation skill.The policy primarily requires instruction-conditioned execution rather than open-ended language generation or autonomous task decomposition.
TurboVLA uses separate visual and lightweight text encoders with direct vision-language interaction to construct control-oriented representations.This V + L → A design replaces the large-language-model-centered execution pathway while retaining language conditioning.
32 Hz: TurboVLA produces action chunks in 31.2 ms on a consumer-grade RTX 4090.The latency is measured end-to-end from receiving the current multimodal observation to producing an action chunk.
0.2B parameters: TurboVLA uses approximately 6% of π0.5’s parameter count while preserving strong manipulation performance.

2 RELATED WORK

Prior VLA research combines visual observations, language instructions, and action prediction using pretrained multimodal representations. Related efficiency efforts redesign action decoding and interfaces, while other work adds geometry-aware or temporal representations for manipulation.

Vision-language-action models: VLA models unify visual observations, task instructions, and action prediction, often leveraging large-scale vision-language pretraining for semantic generalization.RT-1 demonstrated scalable transformer-based robot control; RT-2 and OpenVLA adapted pretrained vision-language models to robot trajectories through action tokens.
Vision-language-action models: Geometry-aware methods use pose-centric pretraining or point-action interaction, while dynamic-manipulation policies incorporate temporal motion cues and short-horizon future prediction.These approaches expand the intermediate representations available for robotic control.
Language interfaces for robot control: Textual instructions can specify tasks and condition perception and control, enabling shared policies to map visual observations and natural-language commands directly to manipulation behaviors.Early imitation-learning methods established this direct language-conditioned control interface.

3 PRELIMINARIES

The preliminaries characterize prevailing VLA models as LLM-centered V → L → A systems, where the language model bridges visual perception and robot action. They contrast this dependency with direct bidirectional vision-language interaction for task-conditioned representations.

LLM-centric vision-language-action models: Most VLA models place a large language model at the center of the vision-to-action pathway, projecting visual features into its token space before processing instructions jointly.The visual encoder extracts features, and a projection maps them into the language-model embedding space.
Direct vision-language interaction: TurboVLA maps visual observations and language instructions to continuous action chunks through compact modality encoders, vision-language interaction, and an action chunk decoder.Its interaction module uses stacked bidirectional cross-attention to produce vision-aware instruction features and instruction-conditioned visual features.
LLM-centric vision-language-action models: The prevailing computation is summarized as V → L → A because the LLM-centered interface aligns visual information with instructions and transforms it into multimodal action-prediction representations.The language model functions as a representational bridge between visual perception and robot action.
LLM-centric vision-language-action models: Autoregressive models discretize and sequentially predict actions, whereas action-expert models use separate continuous decoders to generate actions in parallel.Despite different action-generation mechanisms, both designs retain features produced by the large language model as the decoder input.
Direct vision-language interaction: Cross-attention enables direct exchange between visual and language features, producing language-conditioned visual features and vision-aware instruction features.Bidirectional interaction lets task language shape visual processing while visual context refines the instruction representation.

4 TURBOVLA

TurboVLA implements a direct V + L → A pathway by separately encoding visual observations and instructions, exchanging information through lightweight bidirectional interaction, and decoding continuous action chunks. Robot state enters at action decoding, while compact feature processing avoids an LLM-centered interface.

Architecture: TurboVLA separately encodes visual observations and task instructions, then uses compact vision-language interaction to construct action-ready features for execution-level manipulation.The architecture replaces the conventional LLM-centered interface with a direct V + L → A pathway.
Modality Encoding: Complete instruction-token sequences preserve objects, attributes, and spatial relations for fine-grained visual conditioning rather than pooling them into a single embedding.A lightweight encoder such as BERT processes execution-level instructions, which typically specify manipulation skills without open-ended generation or task-level planning.
Action Decoding: Robot-state features enter directly into a lightweight ACT-style decoder, which predicts continuous action chunks from fused multimodal representations.This keeps cross-modal interaction focused on task-conditioned scene understanding while providing the current embodiment configuration at decoding.
Vision-Language Interaction: Bidirectional cross-modal layers update visual and instruction streams jointly, conditioning visual features on task semantics while adapting instructions to the current scene.The interaction module uses layer normalization, bidirectional cross-attention, modality-specific feedforward networks, and residual connections.
Training and Inference: All H action queries decode in parallel, producing the complete H-step action chunk in one forward pass without tokenization or sequential generation.TurboVLA is trained by behavior cloning on expert action chunks with an ℓ1 loss and no auxiliary language-modeling objective.

5 EXPERIMENTS

TurboVLA is evaluated across single-arm simulation, bimanual simulation, and real-world manipulation, achieving strong task success while substantially reducing inference latency, parameter count, and VRAM. Ablations show that semantic language conditioning and bidirectional vision-language interaction are central to its performance.

Evaluation settings: TurboVLA is evaluated on LIBERO, RoboTwin 2.0, and an AgileX Piper real robotic platform.The benchmarks cover single-arm manipulation, bimanual manipulation, and real-world deployment.
Simulation results: 97.7% average success matches or exceeds larger VLA policies while TurboVLA uses 31.2 ms latency and about 6% of π0.5’s parameters.On LIBERO, TurboVLA achieves 97.7% average success versus 96.9% for π0.5, reducing latency from 93.6 ms to 31.2 ms.
Practical deployability: 0.9 GB of inference VRAM combines with 97.7% average success and 31.2 ms action-chunk inference, enabling practical deployment under memory and latency constraints.TurboVLA also achieves 92.5%, 80%, 90%, and 87.5% success on four real-world AgileX Piper tasks, consistently outperforming π0.5.
Ablations: Removing language lowers average success from 97.7% to 70.8%, including a LIBERO-Goal decline from 97.4% to 11.6%.Replacing semantic instructions with a learned task-ID embedding recovers only part of the performance and remains 2.3% below the semantic-instruction configuration.
Ablations: 95.2% average success from direct concatenation is lower than the proposed bidirectional vision-language interaction under otherwise unchanged settings.The comparison includes no interaction and two asymmetric cross-attention variants before action decoding.
Ablations: N = 6 interaction layers improve average success from 93.5% at N = 2 to 97.7%, whereas N = 8 slightly reduces it to 96.6%.The paper selects N = 6 as a balance between capacity and efficiency.

6 CONCLUSION

TurboVLA proposes a simple, efficient V + L → A paradigm that replaces the conventional LLM-centered pathway with lightweight instruction encoding, compact visual representations, bidirectional interaction, and action-chunk decoding. This design preserves task-conditioned manipulation capability while reducing model size, inference latency, and memory consumption.

Conclusion: TurboVLA introduces a V + L → A paradigm that moves beyond the conventional LLM-centered execution pathway for vision-language-action learning.The approach directly connects vision and language to action rather than centering execution on an LLM.
Conclusion: TurboVLA combines lightweight instruction encoding, compact visual representations, bidirectional vision-language interaction, and action-chunk decoding.These components form the model’s efficient vision-language-action design.
Conclusion: TurboVLA preserves task-conditioned manipulation capability while significantly reducing model size, inference latency, and memory consumption.The conclusion attributes these efficiency gains to the proposed design.