Amap Releases ABot-Recon, a 12-Frame 3D Mapper for 10,000-Frame Video Runs

A robot or mapping device may not need to retain an expanding visual history to build a map over a long route. Amap’s new ABot-Recon instead keeps a rolling 12-frame view, then chains local estimates of geometry and camera movement into a global 3D result across sequences the company says can exceed 10,000 frames.

A short window, composed into a long route

Released August 28, ABot-Recon takes monocular RGB video, meaning video from one camera. Its documented workflow uses features cached from the prior 11 frames, produces a point map for the current frame, and estimates the camera’s position change from one adjacent frame to the next. Those small predictions are composed into a camera trajectory and point cloud for the full sequence.

That design differs from streaming reconstruction systems that retain historical anchors, keyframes, feature caches, or other long-range state to maintain consistency over time. Amap’s stated proposition is that the learned model’s state and work per frame stay independent of how long the sequence has become. The trade-off is clear: repeated local pose errors can accumulate as a route grows.

24.45 FPSReported throughput

Amap reported 24.45 frames per second on KITTI-02; the documented measurement used an Nvidia H100 at 504 × 280 resolution and excluded input storage.

6.71 GiBReported peak memory

The same H100 benchmark reported approximately 6.71 GiB of peak memory.

4.35 mOxford Spires trajectory error

Amap reported average trajectory error of 4.35 meters and relative rotation error of 0.12 degrees on Oxford Spires.

Real time is not yet a consumer-GPU result

The release invokes a GTX 1080 Ti in connection with the pipeline’s memory fit. But the 24.45 FPS and 6.71 GiB figures came from an H100, while the released software environment was validated on an A100. The disclosed evidence therefore does not establish that the older consumer card can deliver the reported speed.

That distinction is operational rather than semantic. Fitting a model in memory and running it in real time are separate thresholds for a system intended to process a live camera feed. Device-specific tests will determine whether ABot-Recon’s fixed-context approach translates to less costly hardware.

The release package is useful, but not fully reproducible

Amap has made the inference implementation, model checkpoint, and evaluation code available. The package includes a Python API, command-line demonstration tools, evaluation instructions, and exports for trajectories and colored global point clouds. It can work from ordinary camera frames without a depth sensor or pre-calibrated camera parameters.

What developers can and cannot do now

• Run and evaluate the released checkpoint and inference implementation.
• Export raw or loop-refined trajectories and colored global point clouds.
• Reproduce training from end to end only after the planned training code and recipes arrive; those materials were scheduled for September 30.
• Use the Apache 2.0 code under its stated terms, while treating the CC BY-NC 4.0 weights as restricted to noncommercial research and education unless separate written authorization is obtained.

A mechanism for managing drift, with one important boundary

Amap documents a motion-visual rotation refiner and a composition-aware pose loss intended to limit accumulated drift. It also provides optional loop closure: a separate backend finds candidate pairs when the camera revisits a place and applies sparse pose-graph refinement. Amap labels its headline Oxford Spires camera-pose result as streaming-only, without loop closure.

The central question is whether those local estimates remain stable beyond the company’s benchmark conditions, especially across cheaper GPUs and difficult routes. ABot-Recon offers a concrete, testable alternative to memory-heavy long-sequence reconstruction. Its practical reach will depend on independent runs, hardware performance, and the eventual availability of the training recipe.