Amap Unveils ABot-Recon for Real-Time 3D Reconstruction of 10,000-Frame Scenes

As autonomous driving and embodied intelligence move into open environments, physical AI needs to continuously understand its spatial context: where it is, what surrounds it, and how to move next. In private scenes like shopping malls, parks, and warehouses, where positioning signals are weak and environments change rapidly, systems cannot rely on one-time mapping or offline reconstruction. Streaming 3D reconstruction has been a mainstream solution for real-time scene building, but it has suffered from error accumulation and excessive GPU memory usage.

To keep long-sequence global consistency, traditional streaming reconstruction methods set up memory anchors and rely on long-term memory to store, retrieve, and fuse historical information. As the sequence lengthens, these methods get slower, less accurate, and more memory-intensive. ABot-Recon instead adopts a “local pose prediction + residual optimization” path. It uses only the latest 12 consecutive frames as a local context window, predicting the local point cloud in the current camera coordinate system and the relative pose between adjacent frames at each step. This keeps computational complexity constant and ensures smooth trajectories, while the global trajectory and 3D scene are assembled online incrementally.

Because the prediction target is always local and independent of sequence length, memory consumption and per-frame computation do not grow with video length. To mitigate error accumulation in local prediction, ABot-Recon integrates correction mechanisms and error constraints in both inference and training, calibrating trajectory errors in real time. As a result, it surpasses existing methods on multiple public long-sequence benchmarks, including KITTI, Oxford Spires, and VBR.

On the Oxford Spires benchmark, ABot-Recon’s average trajectory error is 40.6% lower than that of the representative method LingBot-Map. Its relative rotation error (RPE-R) reaches 0.12°, the best among streaming methods and roughly 40% lower than the previous SOTA. On KITTI-02, it achieves real-time reconstruction at 24.45 frames per second, 1.24 times the inference speed of representative methods. Among all methods with an average trajectory error below 20 meters, ABot-Recon has the highest throughput and the lowest GPU memory usage.

The model’s peak memory footprint is about 6.71 GB, approximately one-third that of similar models, allowing it to run on a consumer-level GTX 1080Ti. This greatly reduces the barrier to streaming 3D reconstruction. ABot-Recon also requires only monocular RGB video input, without extra depth sensors or known camera parameters, making it suitable for surveying and mapping, private-domain mapping, map updating, embodied intelligence, autonomous driving, and 3D content production.

Amap has open-sourced the inference and evaluation code and weights on GitHub and launched an online experience zone on ModelScope for developers. Project page and technical report are available online.