SenseTime SenseAuto presented a new and upgraded “Insight Awareness” world model 2.0 at the World Congress on Artificial Intelligence (WAIC 2025) and released the largest industry-generated driver data set, WorldSim-Drive.
The commercial soup insulation, in close collaboration with the high-end smart car (IM Motors) under the umbrella of the Upper Air Group, has succeeded in connecting data generation to complex driving scenes such as Cut-in (car access) and collisions. In the future, the Business Soup Shadows Project has joined forces with intellectuals and other partners to build a data factory for smart driving, which promotes the scale-up application of generating AI in the field of automatic driving.
The Insight World Model is the core technology for commercial insulation-generated smart driving. In November 2024, the model was first launched on the Shadow Power AI Day, focusing on 2D video generation. The Shanghai Car Show was upgraded to version 2.0 in 2025 to support the near real-time interactive 4D World Model, achieve “4D space free interaction” and significantly increase the efficiency and authenticity of simulation training.
WAIC 2025, “Insight 2.0” is further optimized. Interactive speed is five times higher than the industry ‘ s most advanced level (SOTA), and vehicle decision-making and feedback respond almost instantaneously, optimizing the effectiveness of simulation training. Perception results are 98 per cent consistent with real precision data and effectively avoid the degradation of capacity in Sim-to-Real migration.
Through the VLAR technical architecture, the model is able to replicate 4D for long tailings such as construction blocks, narrow corridors, etc. For example, in the construction of the brakes on the road, R-Uniad relied on the “Insight” to complete the video decomposition, re-engineering and detail restoration, which took hours to produce a fine simulation of the scene, 95 per cent consistent, 1.5 times higher than the traditional 3DGS technology.
The largest industry-generated driver data set, WorldSim-Drive, was released in the WAIC 2025, specifically designed for end-to-end smart driving models. The coverage of diverse scenarios such as urban roads, rural roads, highways, etc. provides high-quality simulation data, especially for the Cut-in, collisions, construction blocks etc. Combined with the 4D simulation capability of “Insomnia 2.0”, WorldSim-Drive generates reality scene data through closed loops of data collection, real value production, model training and real-car testing, addressing bottlenecks of less than 1 per cent in the extraction of effective information on extreme scenes in million-volume car data.
The business insomnia and cooperation with the Zihu auto under the umbrella of the upper-air group are key practices in its production and land. The Zhuihu car is located in a high-end smart electric vehicle brand, and the flag-departure type (e.g. IM L6, LS6) is used to play smart. R-UniAD, a commercial insinuation, has accessed the Cut-in and collision scene data generation links. Through “Insight 2.0”, R-Uniad is conducting 4D simulations of complex scenarios, intensive learning and generalization validations to enhance the model’s ability to respond to long tails.
The World Model 2.0 and WorldSim-Drive data sets presented by the merchant soup dimming on WAIC 2025 mark a major breakthrough in the generation of AI in the field of smart driving. Its collaboration with Sakaki has successfully facilitated the data generation links of the Cut-in, collisions and so on, laying the foundation for a mass landing.