Amap ships ABot-Recon to rebuild 10,000-frame scenes through a 12-frame window

Alibaba's mapping platform Amap released ABot-Recon on August 28, an open-source streaming 3D reconstruction model that rebuilds scenes exceeding 10,000 frames from a fixed 12-frame context, achieving 24.45 FPS on an H100. The model, credited to AMAP CV Lab, is designed for robots and mapping devices with limited memory, though its weights are noncommercial and the reported speed was measured on an H100, not the GTX 1080 Ti mentioned in the release.

Amap ships ABot-Recon to rebuild 10,000-frame scenes through a 12-frame window Alibaba's mapping lab open-sourced inference code and weights, though its 24.45 FPS result came from an H100, not the GTX 1080 Ti invoked in the release. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: PR Newswire https://www.prnewswire.com/news-releases/alibabas-amap-introduces-abot-recon-reconstructing-10-000-frame-scale-3d-scenes-from-just-12-frames-in-real-time-302862527.html Why it matters ABot-Recon turns long-horizon mapping into a fixed-size local prediction task, a promising design for robots with limited memory. Its code is open, but the weights are noncommercial and the reported speed was measured on an H100. Amap https://map.amap.com/about/index.html?ref=runtimewire , Alibaba's mapping platform, released ABot-Recon https://www.prnewswire.com/news-releases/alibabas-amap-introduces-abot-recon-reconstructing-10-000-frame-scale-3d-scenes-from-just-12-frames-in-real-time-302862527.html?ref=runtimewire on August 28, presenting a fixed 12-frame context as an alternative to the growing memory stores used by many long-sequence reconstruction systems. The project page https://amap-cvlab.github.io/ABot-Recon-html/?ref=runtimewire documents the release. The streaming model processes monocular RGB video through a rolling context of 12 consecutive frames. For each new frame, ABot-Recon estimates local geometry and the camera's movement relative to the preceding frame. It then composes those small predictions into a global camera trajectory and point cloud. That bounded design is the point. Many streaming reconstruction systems retain historical anchors, keyframes or feature caches to preserve consistency as a camera moves. Those memories can grow expensive or degrade as a sequence stretches into thousands of frames. ABot-Recon keeps the learned model's state and per-frame work independent of elapsed sequence length, according to the open-source repository https://github.com/amap-cvlab/ABot-Recon?ref=runtimewire . Amap says the result can reconstruct sequences exceeding 10,000 frames while running in real time. The claim is technically consequential for robots, cars and mapping devices that must understand unfamiliar spaces as they move through them. It also comes with a hardware caveat that Amap's release largely smooths over. The lab is the credited author ABot-Recon was released under the collective AMAP CV Lab name. The repository's citation block https://github.com/amap-cvlab/ABot-Recon/blob/main/README.md?ref=runtimewire credits AMAP CV Lab as the author; the supplied materials do not identify an individual project lead or provide a completed paper citation. That collective presentation reflects how Amap is building the broader ABot program inside Alibaba's mapping operation. AMAP CV Lab https://github.com/amap-cvlab?ref=runtimewire describes its work as spanning mapping, autonomous driving, human-centric AI, embodied intelligence, world modeling, 3D generation and reconstruction, and general deep learning. ABot-Recon is a particularly direct expression of Amap's institutional advantage. A mapping platform has practical reasons to care about machines that can estimate where they are, recover surrounding geometry and continue doing both over long routes. The same capability can feed autonomous navigation, private-site mapping and embodied AI training in places where a prepared map is unavailable. The engineering choice is also unusually restrained for a model release. The lab is arguing that a fixed local problem, solved repeatedly, can outperform architectures carrying increasingly elaborate historical state. According to the project's technical README https://github.com/amap-cvlab/ABot-Recon/blob/main/README.md?ref=runtimewire , the model uses KV-cached features from the preceding 11 frames, predicts a point map for the current frame and estimates the relative pose between adjacent frames. The documentation says a motion-visual rotation refiner and a composition-aware pose loss are used to limit the drift that accumulates when thousands of local estimates are chained together. The same README https://github.com/amap-cvlab/ABot-Recon/blob/main/README.md?ref=runtimewire documents optional loop closure: a separate backend retrieves candidate frame pairs and applies sparse pose-graph refinement when the camera revisits an area. Amap labels its headline Oxford Spires camera-pose result as using the streaming model alone, without loop closure. The benchmark headline needs its hardware attached In the August 28 announcement https://www.prnewswire.com/news-releases/alibabas-amap-introduces-abot-recon-reconstructing-10-000-frame-scale-3d-scenes-from-just-12-frames-in-real-time-302862527.html?ref=runtimewire , Amap says ABot-Recon reached 24.45 frames per second on KITTI-02 while using approximately 6.71 GB of peak memory. The release uses that memory figure to argue that a GTX 1080 Ti can run the full pipeline. The repository's benchmark documentation https://github.com/amap-cvlab/ABot-Recon/blob/main/README.md?ref=runtimewire gives the result a narrower definition. The 24.45 FPS and 6.71 GiB measurements were taken at 504 x 280 resolution on an Nvidia H100, with input storage excluded, while the released software environment was validated on an A100. Those are data-center accelerators, so the disclosed evidence does not establish the same speed on the older consumer card named in the release. The distinction matters because "fits in memory" and "runs at the reported real-time speed" are different claims. ABot-Recon may fit on a GTX 1080 Ti, given its 11 GB memory capacity and Amap's stated 6.71 GiB peak. Developers still need a device-specific benchmark before treating 24.45 FPS as a consumer-hardware result. Amap also reports an average trajectory error of 4.35 meters and a relative rotation error of 0.12 degrees on Oxford Spires. According to Amap, the average trajectory error was 40.6% below the previous leading method. The repository's Oxford Spires tables https://github.com/amap-cvlab/ABot-Recon/blob/main/README.md?ref=runtimewire report a 91.81% dense-reconstruction F1 score at a four-meter threshold and specify that the headline camera-pose result excludes loop closure. The figures come from Amap's own technical materials and have not yet arrived in a completed arXiv record with named authors. ABot-Recon enters a fast-moving research contest. LONG3R https://arxiv.org/abs/2507.18255?ref=runtimewire maintains and updates a pruned spatiotemporal memory. LingBot-Map, published as Geometric Context Transformer for Streaming 3D Reconstruction https://arxiv.org/abs/2604.14141?ref=runtimewire , combines anchor context, a pose-reference window and trajectory memory. LongStream https://arxiv.org/abs/2602.13172?ref=runtimewire uses local pose prediction alongside cache refreshes, while HorizonStream https://arxiv.org/abs/2605.23889?ref=runtimewire carries long-range evidence through bounded linear attention. Amap's contribution is a cleaner extreme: keep only the short rolling context inside the learned reconstruction model, then assemble the long route outside it. Open code, restricted weights Amap has released the model checkpoint, inference implementation and evaluation code. The technical README https://github.com/amap-cvlab/ABot-Recon/blob/main/README.md?ref=runtimewire documents a Python API, command-line demo and point-cloud export tools, while a separate evaluation branch https://github.com/amap-cvlab/ABot-Recon/tree/eval?ref=runtimewire provides dataset preparation and benchmark instructions. The README schedules training code and recipes for September 30, so the August 28 package does not yet support complete reproduction from training data. The licensing also draws a line between experimentation and deployment. Amap released the source code under Apache 2.0, while the model weights use a CC BY-NC 4.0 license https://github.com/amap-cvlab/ABot-Recon/blob/main/MODEL LICENSE.md?ref=runtimewire . The weights are designated for noncommercial research and education. Commercial use requires separate written authorization from the relevant rights holders, including any restrictions inherited from the Pi3 weights used to initialize training. That makes ABot-Recon immediately useful as a research artifact and a reference implementation. It is not a drop-in commercial component for a robotics founder building a paid product. The release still gives developers something concrete to test. ABot-Recon accepts ordinary camera frames without a depth sensor or pre-calibrated camera parameters, and its repository documentation https://github.com/amap-cvlab/ABot-Recon?ref=runtimewire covers raw and loop-refined trajectories as well as colored global point-cloud export. Those choices lower the cost of evaluating the central idea: a machine may need only a short visual memory if its local motion estimates are accurate enough to compose into a stable map. Amap has supplied code for that bet. Independent runs across cheaper GPUs, difficult lighting, moving objects and repeated routes will determine how far the 12-frame approach travels outside Amap's benchmark tables.