The Robbyant Team has introduced Ling Bot-Map, a feed-forward 3D foundation model designed for streaming 3D reconstruction. The model features a Geometric Context Transformer, which unifies coordinate grounding, dense geometric cues, and long-range drift correction within a single framework. This architecture utilizes anchor context, a pose-reference window, and trajectory memory for efficient processing.

Ling Bot-Map boasts high-efficiency streaming inference through its feed-forward design and paged KV cache attention. This enables stable inference at approximately 20 FPS for 518x378 resolution over extended sequences, even those exceeding 10,000 frames. The model demonstrates state-of-the-art performance on various benchmarks, outperforming existing streaming and iterative optimization-based methods.

Recent updates include bug fixes for the SDPA KV cache, improved evaluation benchmark scripts for datasets like KITTI and Oxford Spires, and the release of a long-video demo showcasing a 13-minute indoor walkthrough. The team also accelerated the model using compilation and addressed a Flash Infer KV cache bug, leading to better pose and reconstruction quality for sequences longer than 320 frames.