Amap's ABot-Recon builds 10,000-frame scenes from 12

Alibaba’s Amap released ABot-Recon, a streaming 3D reconstruction model that uses just 12 consecutive frames to build scenes spanning more than 10,000 frames in real time. In Amap’s release, it reports 24.45 FPS on KITTI-02, 6.71 GB peak memory, and 40.6% lower trajectory error on Oxford Spires, and says the weights and evaluation code are open-sourced; the project page shows the fixed-window approach. For builders working on world models as robot-simulation infrastructure, the fact to file is the fixed context window: long visual rollouts can be composed without a growing memory bill, but real-world drift and hardware fit still need a local evaluation.