Core Insights - The value of the world model lies in expanding the physical boundaries of AI rather than replacing human cognition, as stated by the CTO of SenseTime, Wang Xiaogang [2] - SenseTime showcased its upgraded "Jueying Awakening" world model at the World Artificial Intelligence Conference (WAIC 2025), emphasizing its ambition to extend into embodied intelligence by constructing a 4D real-world model [2][3] Group 1: Product Development and Capabilities - The "Jueying Awakening" model is the first generative world model in the industry to achieve mass production, demonstrating its technical value in practical applications [3] - SenseTime has collaborated with SAIC's Zhiji Auto to generate critical driving scenarios like Cut-in and collisions, allowing for the batch generation of high-risk, low-probability driving scenarios without relying on real road tests [3] - The newly launched product platform for the generative world model is open for trial use by B-end enterprises and C-end developers [4] Group 2: Data Generation and Efficiency - The platform offers flexible scene customization, allowing adjustments for weather, lighting, and road types, and features convenient prompt word generation for scene video creation [5] - The "WorldSim-Drive" dataset, the largest generative driving dataset in the industry, contains over 1 million clips of production-level data, covering various weather and lighting conditions [5] - The model can generate data equivalent to the collection capacity of 10 real test vehicles or 100 road test vehicles daily, achieving efficiency comparable to 500 production vehicles [5] Group 3: 4D Interactive Training Environment - The 4D interactive training environment integrates 3DGS reconstruction technology with world model generation capabilities, enabling high-precision digital reconstruction of real spaces [6][8] - Users can trigger a closed-loop process to quickly generate complex scenarios through text descriptions or scene layouts [8] - The training environment has been implemented in collaboration with Zhiji Auto, covering typical scenarios and aiming to expand to millions of scenarios to encompass nearly all driving possibilities [8] Group 4: Extension to Embodied Intelligence - SenseTime aims to transfer the "virtual-real fusion" data from the autonomous driving sector to embodied intelligence, addressing the challenges of data dimensional explosion and the Sim2Real gap [10] - The model utilizes multi-modal spatiotemporal alignment capabilities and generates high-fidelity 4D environments to predict object movement trajectories in real-time [10] - The model can generate both first-person and third-person perspectives, enhancing the completeness of data views for robotic applications [11] Conclusion - The evolution of "Jueying Awakening" represents a shift of AI from the digital world to the physical world, with the core value being the transformation of AI creativity into productivity [11]
搭建AI通往真实世界交互的桥梁，商汤“绝影开悟”世界模型再升级