LingBot-Map is a feed-forward 3D foundation model designed for reconstructing scenes from streaming data, developed by the Robbyant Team. The repository has grown modestly since July 2026, gaining 2 stargazers to reach 9,720 total stars. The project is classified across multiple domains including language learning, chatbot functionality, interactive mapping, multilingual support, and educational tools, though the core technical focus is on 3D scene reconstruction from video streams.
The repository's primary innovation is the Geometric Context Transformer architecture, which unifies coordinate grounding, dense geometric cues, and long-range drift correction within a single streaming framework. This is achieved through three key components: anchor context for spatial grounding, a pose-reference window for temporal consistency, and trajectory memory for correcting accumulated drift over long sequences. The architecture enables high-efficiency streaming inference with a feed-forward design that incorporates paged KV cache attention, achieving approximately 20 frames per second on 518x378 resolution video streams that exceed 10,000 frames in length.
The codebase is written in Python and includes comprehensive documentation and multiple demo scripts. The interactive demo uses a browser-based viser viewer accessible at localhost:8080, allowing users to visualize 3D reconstructions in real time. The repository provides four example scenes ready to run out of the box, plus a featured 25,000-frame indoor walkthrough spanning 13 minutes. For long sequences exceeding 3,000 frames, the system supports windowed inference mode with configurable keyframe intervals to manage KV cache memory constraints. The project also includes an offline rendering pipeline for batch processing long sequences.
Installation requires PyTorch 2.8.0 with CUDA 12.8, though the developers note that newer PyTorch versions may work for the interactive demo. The system optionally integrates FlashInfer for optimized paged KV cache attention, with automatic fallback to PyTorch's native SDPA attention if FlashInfer is unavailable. Additional features include sky masking for outdoor scenes using an ONNX segmentation model that automatically downloads on first use, with cached masks to avoid regeneration on subsequent runs.
Model checkpoints are available through both Hugging Face and ModelScope repositories. Three variants are provided: lingbot-map-long optimized for long sequences and large-scale scenes, a balanced lingbot-map checkpoint, and lingbot-map-stage1 for bidirectional inference. The repository includes evaluation benchmarks for multiple datasets including KITTI, Oxford Spires, VBR, Droid-W, TUM-D, 7-scenes, ETH3D, Tanks and Temples, and NRGBD.
Development activity shows median issue and pull request response latency of 3.8 hours across 58 tracked items, with mean latency of 65.5 hours. LinZhuoChen is the most active contributor with 62 tracked events, followed by zhangIU with 8 events and justimyhxu with 4 events. The repository shares contributors with nvidia/tensorrt-llm, colmap/colmap, and rocm/rocm. Recent updates include bug fixes for SDPA and FlashInfer KV cache handling, performance optimizations through PyTorch compilation support, and expanded dataset support. The most actively tracked enhancement label indicates ongoing feature development and improvement efforts.
