Ant Group's LingBot-Map Rebuilds 3D Scenes in Real Time Without Depth Sensors
Ant Group's Robbyant open-sources LingBot-Map, a streaming 3D reconstruction model that runs at 20 FPS over 10,000-frame sequences and beats offline methods on major benchmarks
Takeaways
- What it is: LingBot-Map is an open-source streaming 3D reconstruction model from Ant Group's Robbyant that runs at 20 FPS from plain RGB video.
- Key innovation: Geometric Context Attention (GCA) splits streaming state into three learned memory types — anchor, local window, trajectory memory — reducing per-frame token growth by ~80x vs. causal attention.
- Benchmark dominance: Sets new SOTA across 5 datasets, beating all streaming methods and even offline models like DA3 and VGGT on Oxford Spires large-scale sequences.
- Stability at scale: ATE rises only 0.69 m when sequence length increases 12x (320 to 3,840 frames); competing methods degrade by 10-15 m over the same range.
- Limitations: Fixed 518×378 resolution (~quarter of 1080p), no explicit loop closure, windowed VO mode needed for sequences beyond ~3,000 frames.
- Access: Apache 2.0, 4.63 GB checkpoint on HuggingFace; paper at arXiv:2604.14141.
LingBot-Map is a new open-source 3D foundation model from Robbyant, the embodied AI lab inside Ant Group, that does something the field has been chasing for years: reconstruct a full 3D scene in real time, from a plain RGB camera, as the camera moves. No LiDAR. No depth sensor. No batch processing. Just a continuous video stream in, and a live point cloud with camera poses out.
Streaming 3D reconstruction aims to recover 3D information, such as camera poses and point clouds, from a video stream, which necessitates geometric accuracy, temporal consistency, and computational efficiency. That triple constraint is exactly what has made this problem so hard. Previous approaches either nailed accuracy by processing all frames offline, or achieved streaming but drifted badly over long sequences. LingBot-Map is the first purely autoregressive model to credibly solve all three at once.
The drift problem that killed streaming reconstruction
Most existing 3D foundation models treat reconstruction as a set-to-set problem: you feed in a collection of images, the model reasons about correspondences and geometry, and you get a point cloud out the other end. This works well for offline applications where you have all your data upfront. It falls apart when you need to reconstruct the world as you move through it.
The core failure mode of streaming methods is drift. As the sequence grows, small per-frame pose errors compound, and the estimated trajectory slowly diverges from reality. Existing streaming approaches like CUT3R use a recurrent state that compresses aggressively and forgets geometry. Others like StreamVGGT cache nearly every frame, causing memory and compute to balloon linearly with sequence length. Neither strategy scales.
Instead of accumulating frames and processing them together, LingBot-Map maintains an internal state that represents its current understanding of the scene. Each new frame refines that state rather than requiring a complete recomputation. The key insight is borrowed from classical SLAM systems, but made fully learnable.
Three types of memory, learned end-to-end
A defining aspect of LingBot-Map lies in its carefully designed attention mechanism, which integrates an anchor context, a pose-reference window, and a trajectory memory to address coordinate grounding, dense geometric cues, and long-range drift correction, respectively. Each of these three components maps directly to a problem that classical SLAM solved with hand-crafted heuristics. LingBot-Map replaces all of them with learned attention.

- Anchor Context: The first few frames establish a fixed coordinate system and absolute scale. All subsequent frames attend to these anchors as a stable reference. This solves the scale ambiguity that plagues monocular reconstruction.