LingBot-Map
VerifiedOpen-source model that rebuilds 3D scenes from ordinary video as it streams
Product preview
LingBot-Map
What it is
LingBot-Map is a streaming 3D reconstruction model released by Robbyant, the embodied intelligence company within Ant Group. It takes regular camera footage and, as each frame arrives, predicts the camera position and a depth map, building up a point cloud of the scene without LiDAR. Its Geometric Context Attention keeps a compact memory made of an anchor frame, a local window of recent views and a compressed trajectory history, so per-frame cost stays close to flat on very long videos. The code is on GitHub under the Apache 2.0 license, with weights on Hugging Face and ModelScope, demo scripts and evaluation tooling for researchers and robotics developers.
Key features
- Streaming camera pose estimation
- Per-frame depth prediction
- Long-sequence memory via trajectory tokens
- Interactive and offline rendering demos
What we liked
- + Free and Apache 2.0 licensed
- + Works from a standard RGB camera
- + Handles very long video sequences
Worth noting
- · Needs an NVIDIA GPU and Python setup
- · Aimed at researchers rather than casual users
Editor's notes
More context from AI BirdView
What stands out about LingBot-Map is how it treats 3D reconstruction as a memory problem. Instead of storing every past frame, it keeps a small structured state that the model learns end to end, which lets it keep up with long walkthroughs, drives and aerial footage. Robbyant publishes the model openly, so anyone with a capable GPU can try it on their own video.
- Rebuilds a scene incrementally while video is still playing, rather than waiting for a full batch
- Uses only RGB input, which lowers the hardware bar compared with LiDAR rigs
- Ships with demo scripts, a windowed mode for very long clips and an offline rendering pipeline
- Includes evaluation scripts for public benchmarks such as KITTI and Oxford Spires
- Released under Apache 2.0 with weights on Hugging Face and ModelScope
LingBot-Map is best for robotics engineers, AR developers and computer vision researchers who want a strong open baseline for camera-only mapping. A practical tip: start with the example scenes in the repository to confirm your CUDA and PyTorch versions match the recommended setup, then move on to your own footage and experiment with the keyframe interval to balance speed and quality.
Reader reviews
From people who've used LingBot-Map. Emails are never shown publicly.
No reviews yet. Be the first to share your experience.
Alternatives to LingBot-Map
Other tools we like in this space.
Luma Agents
Luma's agent-driven creative platform with Ray3 video models, formerly Dream Machine
Move AI
Markerless motion capture from ordinary cameras
Immersity AI
Give flat photos and videos real depth motion