TurboVLA is a novel Vision-Language-Action (VLA) model that challenges the prevailing LLM-centric paradigm by replacing the large language model (LLM) core with a direct, lightweight $V+L \to A$ mapping. While existing VLA models (e.g., RT-2, OpenVLA) suffer from significant computational overhead due to the use of billions of parameters in their language-model bottlenecks, TurboVLA achieves high-performance real-time manipulation with only 0.2B parameters, sub-1GB VRAM usage, and a 32 Hz inference rate.

Core Methodology

TurboVLA moves away from the $V \to L \to A$ pathway, where visual observations are projected into a linguistic latent space and processed via an LLM. Instead, it architecture consists of three modular components:

1. Multimodal Feature Encoding: The model uses separate, lightweight encoders for visual observations and linguistic instructions. A visual encoder processes $K$ camera views to extract features $Z_v$, while a BERT-based text encoder processes task instructions into a token sequence $Z_l$. These are projected into a shared latent dimension $d=256$. Robot states $s_n$ are also encoded separately via $f_{state}(s_n)$ to assist in action decoding.

2. Vision-Language Interaction Module: To establish correspondence without an LLM, TurboVLA employs a bidirectional cross-attention mechanism. It iterates through $N$ layers to refine features:

$$(\tilde{V}^\ell_n, \tilde{L}^\ell_n) = \text{FusionLayer}^\ell(V^{\ell-1}_n, L^{\ell-1}_n)$$
This process injects scene context into the instruction stream and conditions visual features on specific task semantics, effectively forming "action-ready" representations ($Z_{vl}$) that isolate relevant objects and spatial relations without requiring open-ended reasoning.

3. Action Chunk Decoder: The fused multimodal features and the robot state are fed into an ACT-style transformer decoder. The decoder uses $H$ learnable action queries to produce a sequence of continuous action chunks in a single forward pass:

$$\hat{A}_n = D_\theta(Q_a, [Z_{vl,n}; Z_{s,n}])$$
This non-autoregressive decoding approach bypasses the sequential token-generation overhead inherent in autoregressive VLA models, enabling high-frequency, responsive execution.

Performance and Efficiency

Experiments on the LIBERO benchmark demonstrate that TurboVLA achieves a 97.7% average success rate, matching significantly larger policies. By eliminating the LLM from the control loop, the model achieves a 31.2 ms inference latency on an NVIDIA RTX 4090. Beyond benchmarks, the model demonstrates scalability to complex, bimanual tasks in RoboTwin 2.0 and robust performance in real-world deployments on an AgileX Piper platform. TurboVLA serves as evidence that execution-level control can be successfully decoupled from generative language backbones, providing a viable path toward deploying sophisticated robotic policies on resource-constrained edge devices.