Alibaba’s Amap Debuts AI Models for Hour-Long Interactive 3D Worlds

Existing generative models typically collapse after roughly a minute of inference, failing to maintain spatial or visual coherence. Amap’s approach addresses this by integrating character control and scene navigation into a unified training objective, utilizing a temporal-consistency algorithm that prevents error accumulation. This architecture lacks a hard upper limit on duration, allowing for extended, stable sessions.

For 3D assets, the ABot-3DWorld-0 model employs 3D Gaussian Splatting to render indoor, street, and aerial environments with physical boundaries and photorealistic textures. A standout feature is the "spatial teleport" mechanism, which allows users to anchor points between generated worlds. This enables a seamless transition between scenes, effectively stitching isolated environments into a continuous, explorable network. These assets are modular, meaning individual scenes can be saved, shared, or repurposed.

By optimizing these tasks to run on consumer hardware, Amap avoids the need for distributed cloud computing. The company has partially open-sourced the model weights and code, providing a test interface via the ABot-World Studio. Potential applications range from training environments for embodied AI to rapid film pre-visualization and immersive digital tourism, marking a shift toward interactive, user-driven exploration.