In the field of embodied intelligence, how to enable robots to perceive physical space as accurately as humans has always been a core challenge. Recently, Robbyant, an embodied intelligence company under Ant Group, officially open-sourced the LingBot-Vision model family. This series of self-supervised vision Transformer models demonstrates outstanding performance in dense spatial perception tasks through an innovative "boundary modeling" approach, even surpassing top models with several times larger parameter scales on multiple metrics.

Current visual foundation models mostly focus on "object recognition," answering what is in the image, but often neglect the most critical elements for physical interaction: object boundaries, contours, and depth information. LingBot-Vision cleverly reverses this priority, taking "boundaries" as native pre-training signals. By introducing mask boundary modeling technology, the model can identify the most information-rich boundary regions in an image and use them as the core of training. This approach not only enables the model to learn semantic understanding but also simultaneously exhibits strong geometric spatial perception capabilities.

image.png

In terms of performance, the flagship model ViT-g/16 of LingBot-Vision has only 1.1 billion parameters, yet it delivered the best results in depth estimation tasks such as NYU-Depth v2. Its performance not only surpasses DINOv3, which has 7 billion parameters, but also uses a training corpus that is only about one-third the size. For practical deployment scenarios with limited computing resources, the series also provides distilled versions ranging from 3 billion parameters to smaller sizes, ensuring leading dense prediction performance across different hardware specifications.

To demonstrate the practical value of this technology, the development team also upgraded the depth completion system LingBot-Depth 2.0. Testing shows that the system significantly improves accuracy when handling transparent objects, a traditional "blind spot" in perception. As the amount of data increases, the performance curve of LingBot-Vision continues to improve without showing the saturation phenomenon commonly seen in traditional models, further proving the great potential of the boundary-centric spatial perception architecture in handling complex real-world environments.

Currently, LingBot-Vision is fully open-sourced on the Hugging Face platform under the Apache-2.0 license, including weights and inference code for four sizes, from giant to small. With the popularization of this technology, developers will be able to endow robots with more sensitive physical perception capabilities at lower computational costs, driving embodied intelligence toward a more precise interactive future.