At the 2026 World Robot Conference Forum, held in Beijing from August 19-22, Wang Xiaogang, Chairman of DaXiao Robotics and co-founder of SenseTime, stated that world models designed for embodied intelligence should possess three core capabilities.

The first is generative intelligence. The most fundamental capability of a world model is to deduce and predict possible future states of the world, which can manifest as video generation or predictions of future latent variables. Producing aesthetically pleasing videos with high-quality pixels alone is insufficient to support robots in interacting with real physical environments.

The second is physical intelligence, which is currently the most urgent shortfall to address. This includes spatial perception capabilities, physical properties of objects, motion laws, and spatial memory.

The third is cognitive intelligence. At this stage, most embodied robots can only complete simple desktop-level operations. To eventually enter households and handle complex scenarios, they need to break down long-chain, multi-step tasks and judge in real time whether actions are executed correctly, which relies on cognitive-level abilities.

Wang noted that video generation, large language models, and multimodal large models have already accumulated a solid technical foundation in generative intelligence and cognitive intelligence, while physical intelligence remains the biggest bottleneck. "To train competent physical intelligence, in addition to video material, it requires support from multi-dimensional data including camera poses, geometric information, physical attributes of objects, and multi-view observations," he said.

All conference transcripts are compiled from on-site shorthand and have not been reviewed by the speakers. This article is republished for the purpose of disseminating more information and does not imply endorsement of the views expressed or verification of the content.