お知らせ • Jul 17
WeRide Unveils WITT, a Physical AI Cognitive Foundation Model Built on Atomic Physical Facts
WeRide unveiled WITT (World Intelligence Toward Truth), a Physical AI Cognitive Foundation Model designed to build AI cognition of the physical world through trusted facts extracted from real-world experience. WITT introduces Atomic Physical Facts (APFs), the smallest verifiable units of information about the physical world, establishing a new fact-based cognitive framework for Physical AI. Built on four core capabilities—Fact Extraction, Fact Reasoning, Fact Verification and Fact Curation—WITT continuously transforms real-world data into trusted learning signals for AI training, evaluation and iteration. Compared with significantly larger general-purpose AI models, WITT reduces token costs by up to 98% and delivers up to 200x greater data-processing efficiency. Leveraging visual-language model (VLM) capabilities, WITT introduces a new concept called Atomic Physical Facts (APFs) and establishes a fact-based cognitive framework for Physical AI. By connecting multimodal information across video, images and text, WITT decomposes continuously evolving real-world environments into verifiable facts that can be identified, reasoned about and validated, establishing a new generation of AI understanding centered on physical facts. WITT stands for World Intelligence Toward Truth and is inspired by the philosopher Ludwig Wittgenstein, whose proposition that “the world is the totality of facts” closely aligns with the underlying logic of Physical AI. To build cognition of the physical world, AI must first identify trusted facts embedded in environments, behaviors, rules, risks and temporal relationships. These facts become the foundation for reasoning, judgment and decision-making. As Physical AI moves from research into real-world deployment, autonomous driving has emerged as the first domain to achieve large-scale commercial validation. Vast amounts of real-world data continue to grow exponentially, identifying and utilizing data with genuine training, evaluation and iteration value remains difficult. High-value long-tail scenarios are inherently scarce, while datasets collected from both L4 autonomous driving operations and production ADAS systems often contain human interventions, inactive segments and other forms of noise. General-purpose AI models can also struggle to interpret complex traffic environments consistently, leading to hallucinations, factual errors and incomplete scene understanding. The industry increasingly needs an efficient and trusted mechanism for understanding data—one capable of continuously extracting meaningful scene facts from real-world driving data, improving the quality and efficiency of training, evaluation and model iteration, and transforming real-world experience into trusted learning signals that drive the evolution of autonomous systems. WITT was developed to address this challenge. Rooted in WeRide’s large-scale autonomous driving operations, WITT continuously extracts patterns, relationships and trusted facts from vast volumes of operational data. Rather than treating data as raw inputs for model training, WITT treats trusted facts as the fundamental building blocks of Physical AI cognition. This foundation enables the model to transform real-world experience into structured knowledge through four core capabilities: Fact Extraction, Fact Reasoning, Fact Verification and Fact Curation. Together, these capabilities create a complete pipeline spanning scene understanding, event attribution, data validation and learning curation—allowing every kilometer of real-world driving data to become a trusted signal for model improvement. WITT identifies and extracts three categories of Atomic Physical Facts from real-world driving data: standard driving facts, multi-agent interaction facts and physically ambiguous conditions. Together, these facts capture everyday traffic behaviors, evolving relationships among traffic participants and uncertainty within complex physical environments. For example, a driving video can be decomposed into multiple Atomic Physical Facts, including reduced visibility caused by rain, a pedestrian entering a crosswalk, an ego vehicle slowing down, a nearby vehicle traveling in parallel, changing traffic signals and increasing collision risk. Each fact is designed to be highly reliable, traceable and verifiable, enabling richer scene descriptions and providing the foundation for subsequent reasoning, validation and learning. After extracting facts, WITT analyzes key events, behavioral relationships and evolving risks within a scene, while identifying the underlying causes and potential trajectories of those events. Powered by an integrated video intelligence engine, WITT enables users to retrieve target scenarios through keywords or natural-language queries, dramatically improving the efficiency of scenario discovery, data investigation and root-cause analysis. To reduce hallucinations commonly associated with general-purpose AI models, WITT evaluates outputs across six dimensions: Vulnerable road users, Ego-vehicle behavior, Surrounding vehicle behavior, Scene understanding, Comprehensive fact, Traffic facilities. The model introduces factual confidence scoring and validates conclusions against external physical evidence to determine whether interpretations are supported by observable reality. By tracking factual errors, hallucinations, omissions and temporal inconsistencies, WITT provides both a quality benchmark for data users and a preference signal for model training, continuously guiding AI systems toward more accurate and physically grounded understanding. WITT achieves an average factual error rate approximately one-third that of leading general-purpose AI models in autonomous driving scenario understanding tasks. WITT automatically identifies high-value facts and routes them into the most effective learning workflows to maximize model improvement. Rare long-tail scenarios can be returned to WeRide GENESIS, the company’s proprietary general-purpose simulation model, for simulation training and scenario expansion. High-frequency everyday scenarios can support reinforcement learning and workflow optimization. Abnormal or ambiguous data can be directed into review processes to prevent valuable information from being mistakenly discarded as noise.