At WAIC 2026, China's embodied-AI race moves to models and benchmarks

At WAIC 2026 in Shanghai, ACE Robotics launched Kairos 3.1 on July 19, an 'action-oriented' embodied world model, alongside a new industry benchmark called PHYSICAL IQ — a unified test for embodied physical intelligence initiated with the Shanghai AI Association, the Shenzhen Loop Area Institute and the East China branch of CAICT, with more than 20 university and industry partners. Chairman Wang Xiaogang argued that errors in the physical world carry consequences digital models don't, a case for benchmarking real-world execution rather than just understanding.

A day later, Beijing-based ModelBest unveiled its MiniCPM-Robot series of embodied-AI edge models, built to run on the robot rather than in the cloud. Demonstrations included an offline venue-guide agent built with Leju Robotics, doing local question-answering and obstacle navigation with no network. ModelBest said its in-cabin SuperMate product already ships in mass-production vehicles including the Changan Mazda EZ-60 and Geely Galaxy M9, and that its edge models will come pre-installed on several Samsung flagship devices.

The pattern is a maturing stack. China's embodied-AI story has been about hardware and deployment volume; WAIC 2026 showed the software layer filling in — world models to predict, edge models to run locally, and a shared benchmark to compare. Benchmarks and demos are not the same as reliable products, and these launches are vendor-framed. But a common yardstick like PHYSICAL IQ, if adopted, is the kind of unglamorous infrastructure that lets a field move from claims to comparisons.