Robbyant / lingbot-map

2026-04-15 · Source: Github Trending: All languages · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

LingBot-Map is a new feed-forward 3D foundation model designed for high-efficiency streaming 3D reconstruction. Developed by the Robbyant Team, it features a Geometric Context Transformer that unifies coordinate grounding, dense geometric cues, and long-range drift correction within a single streaming framework. The model achieves stable inference at approximately 20 FPS on 518x378 resolution over long sequences exceeding 10,000 frames, utilizing a paged KV cache attention architecture. It demonstrates superior performance on diverse benchmarks compared to existing streaming and iterative optimization methods. The project provides comprehensive tools for installation, interactive demos, and an offline rendering pipeline for batch processing of very long videos, supporting PyTorch 2.8.0 with CUDA 12.8 and FlashInfer for optimal performance.

Key takeaway

For Machine Learning Engineers developing real-time 3D reconstruction systems, LingBot-Map provides a robust, high-performance solution. You should consider integrating this feed-forward model to achieve stable 20 FPS inference on long sequences, even with limited GPU memory by offloading to CPU. Utilize its keyframe interval and windowed inference modes to manage very long video streams effectively, and customize rendering pipelines with YAML for specific visualization needs.

Key insights

LingBot-Map offers a feed-forward 3D foundation model for high-efficiency, state-of-the-art streaming 3D reconstruction.

Principles

• Unify geometric cues and drift correction.
• Paged KV cache enables long-sequence streaming.
• Optimize inference for real-time performance.

Method

LingBot-Map employs a Geometric Context Transformer with anchor context, pose-reference window, and trajectory memory for unified streaming 3D reconstruction. It uses a feed-forward architecture with paged KV cache attention.

In practice

• Use "--keyframe_interval" for long sequences.
• Employ "--offload_to_cpu" for limited VRAM.
• Configure camera paths via YAML for offline rendering.

Topics

• 3D Reconstruction
• Geometric Context Transformer
• Streaming Inference
• Transformer Models
• Real-time Systems
• Computer Vision
• SLAM

Code references

• Robbyant/lingbot-map
• nerfstudio-project/viser
• ureeey/lingbot-map-rtx4060-8g
• NVIDIAGameWorks/kaolin
• facebookresearch/vggt

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• LongStream: Long-Sequence Streaming Autoregressive Visual Geometry — LongStream enables stable, metric-scale 3D reconstruction for ultra-long sequences by decoupling gauge and optimizing Transformer caches.
• Research, easy to read and understand by flookkrup EP.3: — LingBot-Map uses selective memory and contextual anchoring to achieve robust, scalable Streaming 3D Reconstruction without memory overload or drift.
• Toward Semantic Communication for Real-time Mobile 3D Reconstruction — Integrating pixel-wise confidence maps into semantic communication enhances real-time 3D reconstruction robustness.
• Next AI Leap: 3D Memory — Zip-Map enables linear-time 3D reconstruction by embedding a dynamically trained MLP within a frozen transformer.
• WAT3R: Feedforward Underwater 3D Reconstruction — WAT3R is a feed-forward framework that uses a neural adaptation module to improve underwater 3D reconstruction by accounting for imaging degradation.
• Stream3D: Sequential Multi-View 3D Generation via Evidential Memory — Stream3D uses an evidential memory and evidence score mechanism to enable consistent 3D generation from sequential multi-view inputs without retraining existing models.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Github Trending: All languages.