Amap, Alibaba’s location-services platform, released ABot-Recon on August 28, 2026. The model is a streaming 3D reconstruction system that, Amap says, needs only 12 consecutive frames to rebuild scenes that stretch past 10,000 frames in real time. The company published inference code, evaluation scripts, and pre-trained weights on GitHub.

Amap argues that older streaming systems keep memory anchors so they can fuse history and stay consistent. As the video gets longer, those systems slow down, drift, and eat more memory. ABot-Recon instead stays inside a fixed 12-frame window. It predicts a local point cloud and the relative pose between adjacent frames. An online composition step then stitches the global trajectory as it goes, so compute stays roughly constant as the sequence grows.

Benchmarks on a consumer GPU

The company reports state-of-the-art scores on KITTI, Oxford Spires, and VBR. On Oxford Spires, it says average trajectory error fell 40.6 percent versus the prior leading method, with a relative rotation error of 0.12 degrees, about 40 percent lower than the previous mark. On KITTI-02, Amap says the model reconstructs at 24.45 frames per second, 1.24 times faster than existing approaches, with peak memory of about 6.71 gigabytes. That, it claims, is enough to run the full pipeline on a consumer GTX 1080 Ti.

Peak memory is about one-third that of comparable methods, according to the release. Input is monocular RGB video. Amap says the model does not need depth sensors or pre-calibrated camera parameters. It lists private-area mapping, embodied-AI training, autonomous driving, and 3D content production as the jobs where pre-built maps are missing and a live reconstruction has to keep up.

Decoded Take

Amap is not launching a consumer feature. It is publishing a spatial model that tries to make long-drive reconstruction cheap enough for a used gaming GPU. If the GitHub weights reproduce the KITTI and Oxford numbers outside Amap’s harness, this becomes a useful open baseline for teams that cannot afford a lidar stack or a data-center renderer. If they do not, it stays a well-lit press chart. The next check is whether robotics and driving labs actually fork the repo, or whether ABot-Recon remains a research page attached to a maps company.