TurboVLA: a real-time VLA model that runs under 1GB VRAM

Introduction

Vision-language-action models have become a central direction in embodied AI, but many of them inherit a costly design pattern: visual observations are projected into the representational space of a large language model, and actions are decoded after this LLM-centered mediation. This V→L→A route can be powerful, yet it also means that every policy call pays the computational and memory price of passing through a large central interface.

TurboVLA proposes a simpler alternative. Rather than treating the language model as the core bridge between perception and control, it reformulates the problem as a direct V+L→A mapping: visual observations and language instructions are encoded separately, exchanged through lightweight bidirectional vision-language interaction, and then used to predict continuous chunks of robot actions.

Key points

• A different VLA pathway: TurboVLA moves away from the common V→L→A formulation and instead builds task-conditioned action representations directly from visual and linguistic features.
• Compact architecture: The model uses independent encoders for vision and language, a lightweight interaction mechanism between the two modalities, and a compact action decoder.
• Real-time inference claim: On a consumer-grade RTX 4090, the paper reports 31.2ms inference latency, corresponding to about 32Hz control frequency.
• Low memory footprint: TurboVLA uses only 0.9GB inference VRAM and has 0.2B parameters, according to the reported results.
• Strong benchmark performance: On LIBERO, it achieves 97.7% average success and is described as matching or outperforming substantially larger VLA policies.
• Open code: The authors provide a public GitHub repository, making the approach easier to inspect and reproduce.

Why it matters

The main contribution of TurboVLA is not merely that it is smaller. It questions whether a large language model must always serve as the central hub in robotic policy inference. For many manipulation tasks, the essential requirement is to condition action generation on the current visual state and a language instruction. A specialized architecture that lets these two signals interact directly may be more efficient than routing them through a large LLM representation space.

This is especially relevant for robotics deployment. Robot control loops often demand low latency, stable inference, and modest hardware requirements. If a VLA policy can run at real-time rates on a single consumer GPU while using less than 1GB of inference memory, it lowers the practical barrier for experimentation and downstream applications.

The evidence in the provided material is centered on LIBERO, so broader validation on real robots, long-horizon tasks, and more diverse environments will still be important. Even so, TurboVLA offers a clear architectural message: efficient embodied policies may not need to scale by placing ever larger language models at the center. Direct, lightweight vision-language-to-action designs could become an important branch of VLA research.

Source: Hugging Face Daily Papers