New Model Enhances Real-Time Spatial Understanding
On April 16, Robbyant, the AI arm of Ant Group, announced the launch of LingBot-Map, a groundbreaking open-source streaming 3D reconstruction model. This model empowers robots, autonomous vehicles, and AR devices to perceive and understand their environments in real-time using a standard RGB camera.
The new technology operates on a 'see-as-you-go' principle, continuously estimating the camera's position and reconstructing the 3D structure of the scene as video is captured. Unlike traditional methods that process images offline, LingBot-Map provides immediate results, enhancing its utility in fast-paced environments.
LingBot-Map has set a new standard in accuracy. For instance, on the Oxford Spires dataset, known for its challenging lighting, the model achieved an Absolute Trajectory Error (ATE) of just 6.42 meters. This figure represents a near 2.8x improvement in trajectory accuracy over previous methods and outperforms offline models like DA3 and VIPE significantly.
LingBot-Map's Technological Achievements
LingBot-Map excels across other benchmarks such as ETH3D, 7-Scenes, and Tanks and Temples. On the ETH3D benchmark, it achieved a reconstruction F1 score of 98.98, surpassing the second-best method by over 21 per cent. This demonstrates the model's leadership in both pose estimation and 3D reconstruction quality.
In terms of performance, LingBot-Map supports real-time applications with an inference speed of approximately 20 frames per second. It can handle long video sequences exceeding 10,000 frames without compromising accuracy, which is crucial for continuous spatial awareness applications such as robot navigation and obstacle avoidance.
The core of LingBot-Map's innovation is its Geometric Context Attention (GCA) mechanism. This feature efficiently manages geometric information across frames, maintaining essential historical context while minimizing redundant computations. Inspired by classic SLAM systems, the architecture leverages a unified model to handle complex tasks typically requiring intricate designs.
Robbyant continues to expand its open-source suite with models like LingBot-Depth, LingBot-VLA, LingBot-World, and LingBot-VA, further enhancing its technology stack for real-time spatial understanding and mapping.
For more information about LingBot-Map, interested parties are encouraged to visit Robbyant's GitHub page or access their technical report on arXiv. Further applications and details are also available on Robbyant's website. the full announcement can be viewed at Business Wire’s website, providing a comprehensive overview of the model’s capabilities and potential applications.
Last updated: 16 April 2026, 11:34 pm
