WeRide Unveils WITT Physical AI Cognitive Foundation Model with Atomic Physical Facts
WeRide introduced WITT, a Physical AI Cognitive Foundation Model that extracts trusted facts from real-world driving data to improve AI training, reduce token costs by 98%, and boost efficiency by 200x.
98%
200x
10,000 minutes
What Happened
WeRide has unveiled WITT (World Intelligence Toward Truth), a Physical AI Cognitive Foundation Model that builds AI cognition of the physical world through trusted facts extracted from real-world experience. Inspired by philosopher Ludwig Wittgenstein's idea that 'the world is the totality of facts,' WITT introduces Atomic Physical Facts (APFs) as the smallest verifiable units of information. The model uses visual-language model capabilities to decompose real-world environments into verifiable facts, enabling more accurate reasoning and decision-making for autonomous driving.
98%
Compared with significantly larger general-purpose AI models
200x
In comparable workloads versus general-purpose models
- Fact Extraction: Identifies standard driving facts, multi-agent interactions, and physically ambiguous conditions.
- Fact Reasoning: Analyzes events, behavioral relationships, and risks; enables natural-language search for long-tail scenarios.
- Fact Verification: Evaluates outputs across six dimensions with factual confidence scoring; achieves average factual error rate about one-third that of leading general-purpose AI models.
- Fact Curation: Routes high-value facts to learning workflows (e.g., rare long-tail scenarios to GENESIS simulation, common scenarios to reinforcement learning).
Within WeRide’s Physical AI architecture, WITT serves as the understanding and evaluation layer, working with the GENESIS simulation model to form a Physical AI flywheel. WITT extracts facts from real-world data, GENESIS generates high-fidelity simulations and long-tail scenarios, and together they continuously improve vehicle-side models. This flywheel has enabled WeRide to become the world’s only company with large-scale commercial deployment of both L4 autonomous driving and L2++ intelligent driving systems.
WeRide has obtained autonomous driving permits in eight countries and operates over 3,000 autonomous vehicles across 40+ cities. Its L4 Robotaxi services run fully driverless in Guangzhou, Beijing, Abu Dhabi and Dubai. The L2++ solution WRD 3.0 has won six consecutive China Urban Intelligent Driving Competitions and been selected for nearly 30 vehicle programs, including models from Chery Exeed and GAC Aion, with testing expansion into Germany, France and Japan.
Previously from WeRide Inc.
WeRide recently began on-road testing and localization validation of its L2++ end-to-end intelligent driving solution in Germany, France, and Japan. Its GENESIS world model won the 'Generative AI Platform of the Year' award at the 2026 AI Breakthrough Awards. Additionally, the company plans to launch commercial robotaxi services in Zurich later this year in partnership with Uber, pending regulatory approval.
- WeRide Begins On-Road Testing of L2++ Intelligent Driving Solution in Germany, France, Japan
- WeRide's GENESIS world model wins 'Generative AI Platform of the Year' award
- WeRide and Uber to Launch Robotaxi Service in Zurich Later This Year
Background drawn from MotorClaw's earlier coverage of WeRide Inc.'s official releases.
Why this matters
WITT establishes a fact-based cognitive framework for Physical AI, addressing hallucinations and data inefficiency. By reducing token costs and processing vast video data on a single GPU, it enables more reliable autonomous driving systems. This affects AI developers and autonomous vehicle companies seeking to scale L4 and L2++ deployments.
Terms in This Story
- Atomic Physical Facts (APFs)
- The smallest verifiable units of information about the physical world, used to build AI cognition.
- Physical AI
- AI systems that perceive, understand, and interact with the physical world, such as autonomous vehicles.
- Visual-Language Model (VLM)
- An AI model that processes and connects multimodal information like video, images, and text.
Summarised from the linked release; details can be imperfect — always verify against the original source.