Nearly 200 builders gathered at our Spatial Intelligence & World Models event @CVPR 2026 in Denver. The community is hungry for real answers on robotics, world models, video gen, and physical AI. No settled path yet. That’s what makes it interesting. Here are 6 sentences that captured the conversation. 🧵

“Good robot data stays stuck – it cannot be shared across different robots, and it dies when the hardware evolves.” — Aurora Feng, Neural Motion #Robotics #DataFlywheel

“One model (NM-GenET) can work as a data layer that transforms source embodiment data into target-ready data, or as a foundation policy model built directly on its architecture.” — Haoyi Niu, Neural Motion #CrossEmbodiment #FoundationPolicy

“When you train a world model strong enough to model every possible next state, it naturally becomes a strong policy.” — Max Li, NVIDIA Cosmos #WorldModel #PhysicalAI 📷

Simulation may be the only scalable platform for centralized robot evaluation, but it has to solve an impossible triangle: physical and visual fidelity, task diversity, and iteration efficiency. — Jie Wang University of Pennsylvania, GRASP Lab.

“Instead of fixed time chunks, robots need event chunks that align vision and language in meaning, and vision and action in time – improving data efficiency by 5x.” — Xiaofan Li, X Square Robot #EventCentric #WALLWM

“Before a robot executes an action, verify if the imagined future is geometrically consistent. A broken future means a broken action.” — Zesen Zhao, University of Michigan #GeometricReasoning #WorldActionModel

“The next frontier for video models is RL tuning with rich explanation feedback from VLMs – turning frozen gradients into precise, unhackable token-level rewards.” — Gordon Qian, Snap #RLTuning #VideoDiffusion