WeRide Releases WITT Physical AI Foundation Model for Autonomous Driving

WeRide has released WITT, a physical AI foundation model designed to serve as the perceptual and cognitive backbone for the company's autonomous driving systems and broader robotics applications³.

The model uses what WeRide calls "minimum physical fact units" as its fundamental building blocks for unified multimodal scene understanding. The concept draws inspiration from how humans understand the physical world through discrete, composable observations. Each minimum physical fact unit represents a quantum of physical understanding that can be combined and composed to represent arbitrarily complex scenes.

Rather than processing raw sensor streams through separate perception modules for detection, tracking, prediction, and planning, WITT unifies these traditionally separate functions within a single framework that reasons about the physical world at a fundamental level. The company has described this as a bet that physical AI requires fundamentally different architectural approaches from language-only or vision-only models.

The model is designed to handle the diversity of autonomous driving scenarios that traditional modular approaches struggle with. In complex urban environments with unpredictable actors, occluded viewpoints, and varying weather conditions, WITT can leverage its unified representation to maintain consistent scene understanding across different sensor modalities and operating conditions without requiring separate model architectures for each task.

WeRide was founded in 2017 by former Baidu autonomous driving executives. The company has robotaxi operations in multiple Chinese cities and in the UAE and Singapore. WeRide went public on Nasdaq in 2024.

ANALYSIS The architectural choice to collapse traditionally separate perception, tracking, prediction, and planning modules into a single unified framework marks a departure from the modular stack design common in autonomous driving systems. Whether WITT's "minimum physical fact unit" abstraction delivers practical gains over conventional approaches will depend on benchmark and real-world deployment results that WeRide has not yet detailed in the available materials.