Alibaba's Amap launches ABot-Recon for real-time 3D reconstruction

Amap, the location-based services platform owned by Alibaba, has released ABot-Recon, a new streaming 3D reconstruction model. The technology is designed to reconstruct large-scale 3D scenes, spanning over 10,000 frames, using only 12 consecutive frames in real time.

Unlike conventional systems that rely on long-range memory anchors—which can become slower and more memory-intensive as sequences grow—ABot-Recon utilizes a fixed 12-frame local context window. It predicts local point clouds and relative poses between adjacent frames, using an online composition mechanism to assemble a global trajectory incrementally. This approach maintains constant computational complexity regardless of the sequence length.

The model aims to support autonomous driving and embodied AI systems by providing continuous spatial understanding. ABot-Recon reportedly reduces peak memory usage to approximately one-third of comparable methods, making real-time reconstruction possible on consumer-grade hardware. In benchmark testing, such as the Oxford Spires long-sequence benchmark, the model reduced average trajectory error by 40.6% compared to previous leading methods.