Westlake University Opens a World Model Research Center as Its Startup Awomo Claims Four Robotics Benchmark Wins

Westlake University, the young research university in Hangzhou founded by Shi Yisong's generation of returnee scientists, formally launched its Westlake World Model Center at the Yungu campus on October 31, positioning the school in the same race that has pulled Fei-Fei Li and Yann LeCun into world-model startups and prompted Google DeepMind's Genie 4 announcement. Around 90 industry R&D executives and investors attended the founding event and its "Physical World Model Frontier Forum."

The center is led by professor Yu Kaicheng, 33, who runs Westlake's AutoLab, with academician Jin Yaochu serving as chief scientist. Yu's path is a compressed history of the field: first-class honors at the University of Hong Kong, a PhD at EPFL in automatic machine learning, then stints at Intel's smart-systems lab, Abacus.AI and Alibaba's DAMO academy working on autonomous driving, before joining Westlake in 2023 to start AutoLab — when world models were still a niche term.

The lab's history includes an instructive self-reversal. Its first route — generating synthetic extreme driving scenarios to train autonomy models — won the 2024 ECCV autonomous driving track by a wide margin. But after deploying with a major Chinese automaker, Yu's team concluded that if a model could already generate photorealistic data, it had in some sense understood the world, and pixel-perfect rendering was the wrong target. At the end of 2024 AutoLab pivoted to "concept world models": instead of generating pixels, the model decomposes scenes into reusable physical concepts — spatial relations, materials, contact states — in an abstract latent space, and reasons by recombining them. The team reports significantly higher success rates than imitation-learning and reinforcement-learning baselines on complex tasks, at lower prediction cost.

The commercial vehicle for that bet is Awomo, a physical-world foundation model company founded in early 2026. The startup says it has raised more than 100 million RMB in cumulative seed and angel funding from Innoangel Fund, Dongfang Jiafu, Zhengxuan Investment, Tianqi Capital, Westlake Innovation Fund and Jinma Investment. Awomo claims first place on four international robotics benchmark suites covering environmental disturbance, fine manipulation and long-horizon tasks — all vendor-reported results, and no third-party reproduction has been published yet.

The economic argument Yu makes is the same one driving the world-model camp broadly: if a general physical foundation model exists, downstream developers should not need to collect millions of hours of task data; 20 to 30 hours of post-training could suffice to specialize a model for a domain. That claim remains unproven at scale, but it explains why investors crowded into the Yungu campus auditorium — the pitch is a potential shortcut around embodied AI's data bottleneck.

For now, the center is a signal as much as an institution: Chinese academia is organizing around physical AI the way it organized around LLMs three years ago. Whether concept world models generalize beyond automotive and manipulation benchmarks will decide whether Awomo's four benchmark wins were a starting line or a high-water mark.