XPeng Completes the Visual Foundation of Physical AI

On July 21, XPeng Group announced the official release of TuringViT, a high-efficiency visual encoder designed for Vision-Language Model (VLM) and Vision-Language-Action (VLA) application scenarios. The company has systematically reconstructed the architecture design, data paradigms, and training processes. XPeng stated that this encoder will be deployed across three major business scenarios: intelligent driving, smart cockpits, and the IRON humanoid robot.

In VLM/VLA systems, visual encoders are typically regarded as the "perception gateway" of physical AI, responsible for converting images or videos into visual features for subsequent processing by language and action modules.

As VLM/VLA technology moves toward real-world implementation, high-resolution images, multi-view inputs, and continuous video frames impose higher demands on the efficiency and temporal processing capabilities of visual encoders.

XPeng believes that the industry's common approach of "reusing open-source general-purpose ViTs" struggles to simultaneously meet the requirements for performance, latency, and customization in scenarios such as intelligent driving and embodied robotics.

XPeng summarizes the design of TuringViT into three dimensions.

In terms of architecture, TuringViT primarily utilizes Turing Linear Attention (TLA) while retaining a small amount of standard multi-head attention, forming a hybrid Turing Block architecture. Under high-resolution inputs, computational complexity shifts from quadratic scaling to near-linear.

According to test data released by XPeng, at a resolution of 1536×1536, the inference throughput of TuringViT-18L is 3.04 times that of Seed1.5-ViT.

The encoder is available in two specifications: TuringViT-18L, which contains three sets of Turing Blocks and is oriented toward deployment under dynamic resolutions; and TuringViT-24L, which contains four sets of Turing Blocks and focuses on enhancing representation capabilities.

Regarding data, TuringViT adopts the VISTA-Curation multimodal data governance pipeline. According to XPeng, this pipeline performs multi-stage filtering, re-captioning, and scoring of image-text and video data to increase the amount of supervised information per sample. TuringViT was pre-trained on 0.85 billion image-text pairs, approximately 10% of the training data scale used for SigLIP2-L. On six zero-shot classification benchmarks, including ImageNet-1K, TuringViT-24L achieved an average accuracy of 83.6%.

In terms of training, TuringViT employs a four-stage progressive native dynamic resolution training scheme. It adapts to the input characteristics of downstream VLM/VLA models from the pre-training stage and, combined with 2D rotary positional encoding, supports inputs of various sizes and aspect ratios. XPeng stated that this design reduces reliance on the path of "fixed-resolution pre-training + post-hoc adaptation."

XPeng stated that TuringViT will be applied in three types of business scenarios.

In the field of intelligent driving, XPeng describes TuringViT as the core visual encoder for its second-generation VLA model. It is responsible for processing high-resolution images from multiple surround-view cameras and multi-frame dynamic road scene inputs, providing visual tokens for predictive world models.

In smart cockpit scenarios, the native VLM characteristics of TuringViT are used to align visual features with language models and support visual inputs of different aspect ratios and proportions.

Within the technical framework of XPeng's IRON humanoid robot, TuringViT is positioned as the foundational visual module, undertaking functions such as object recognition, spatial relationship understanding, manipulable area detection, and dynamic environment tracking.

At the earnings conference on March 20, He Xiaopeng stated that mass production of the XPeng IRON humanoid robot is planned for the end of 2026, with a monthly capacity target of thousands of units by year-end, and it will be prioritized for commercial deployment in XPeng stores.

XPeng indicated that the architectural design and training process of TuringViT do not depend on specific hardware platforms, offering the industry a reproducible pathway for training large visual models.

From a technological layout perspective, TuringViT completes the visual encoder component within XPeng's physical AI technology system.

Previously, XPeng had sequentially disclosed its multi-view generative world model X-World, the world model inference acceleration engine X-Cache, the predictive world model X-Foresight, and the X-Mind technical framework. TuringViT sits at the perception input end of this physical AI technology system.

However, the transition of visual encoders from technical release to scaled implementation in actual scenarios still requires engineering adaptation, computing power deployment, and scenario validation.

Whether TuringViT can operate stably across the three differentiated physical environments of vehicles, cockpits, and robots remains to be seen. XPeng stated that it will continue to expand the scale of high-quality image-text-video data and deepen technical exploration in temporal modeling and embodied vision.