LingBot-Map Ships Feed-Forward Streaming 3D Reconstruction at 20 FPS — Ant Group's Geometric Context Transformer Handles 10K+ Frame Sequences Without Optimization

arXiv·medium signal

LingBot-Map from Ant Group's Robbyant team introduces a feed-forward 3D foundation model that reconstructs scenes from video streams at 20 FPS on 518x378 resolution using a Geometric Context Transformer architecture. Handles sequences exceeding 10,000 frames with no test-time training or post-optimization, using anchor context for coordinate grounding, a pose-reference window for local geometry, and trajectory memory for global consistency. Achieves state-of-the-art on Oxford Spires, 7-Scenes, Tanks and Temples, and ETH3D benchmarks.