Pudu Robotics Unveils 'One Brain, Multiple Forms' Strategy at WAIC with PuduFM Foundation Model and PuduAgent OS

The embodied intelligence industry is becoming increasingly pragmatic. Rather than building ever more human-like robots, the focus has shifted to which companies can keep robots running reliably in real-world environments while accumulating operational experience. At the 2026 World AI Conference (WAIC) in Shanghai, Pudu Robotics — now the global leader in commercial service robots by market share — laid out its answer.

A newly released independent market report by Frost & Sullivan ranked Pudu Robotics first worldwide in four dimensions: global commercial service robot revenue, global commercial service robot shipments, global commercial cleaning robot revenue, and Chinese commercial service robot overseas revenue. The company's products have now been deployed in 85 countries and regions, surpassing 130,000 units across 16 industries including food service, hospitality, industrial facilities, and warehousing.

At WAIC, Pudu unveiled its top-tier technical strategy — "One Brain, Multiple Forms." The core idea is to let robots of different form factors share a single, continuously evolving intelligent brain rather than developing intelligence independently for each hardware variant. Two products anchor this strategy: the PuduFM embodied intelligence foundation model and the PuduAgent embodied intelligence operating system.

PuduFM serves as the unified physical intelligence brain. It consists of three components. The Physical Intuition Model (PIM) uses a causal-attention Transformer architecture to distill physics dynamics — gravity, friction, center-of-mass shifts — into machine-readable features, enabling action pre-judgment before contact. The VLA multimodal alignment module merges vision, speech, navigation, and manipulation into one reasoning framework, bridging the gap between reaching a location and performing a task. The World Model simulation engine, powered by a Diffusion Transformer architecture, uses real-world data accumulated across 16 industries to generate diverse training trajectories including adversarial edge cases like collisions and slippage.

PuduAgent is the operating system layer that packages PuduFM's capabilities for real-world deployment. It includes three modules: Agent Core for environment understanding and task planning, Agent Memory for persistent task recall, and Agent Runner for real-time execution. The PuduAgent Skills library abstracts navigation, grasping, human-robot interaction, obstacle avoidance, and multi-robot coordination into standardized, reusable atomic skills. PuduAgent Safety acts as a safety guardrail, pre-evaluating action feasibility before execution and intervening under anomalous conditions.

The PUDU D7 humanoid robot, making its debut at WAIC, is the first major hardware showcase for PuduFM. Built for industrial and retail settings, it demonstrates seamless transitions between long-distance navigation and close-range manipulation. Pudu disclosed that its fleet generates 36.5 million hours of navigation data and 15.8 million hours of operational data annually — first-person data from real environments that feeds a flywheel of expanding deployment, data return, model improvement, and broader application coverage.

Notably, Pudu proposed a formal definition: Physical Agent = Hardware Body + System + Skills. This layered decomposition — body, orchestration system, and capability — offers a modular reference architecture for the industry. Frost & Sullivan's report also observes that the robotics industry is evolving from "one robot, one brain" toward "one brain, multiple forms," with Pudu among the earliest adopters of this paradigm shift.

As embodied intelligence enters the deep-water zone of commercialization, the ability to accumulate real-world experience at scale and share intelligence across hardware form factors will likely determine the next phase of the robotics industry. Pudu's WAIC showcase signals that Chinese service robot manufacturers are transitioning from hardware vendors to intelligence platform providers — a shift worth watching closely.