Ant Group open-sources LingBot-Vision: a 1B model for robot spatial perception

Ant Group's robotics division Robbyant published on July 7, 2026, in open access LingBot-Vision — a family of self-supervised Vision Transformer models with 1 billion parameters for dense spatial perception tasks. Along with the code, the team presented the masked boundary modeling approach, which makes object boundaries a native training signal.

What is masked boundary modeling

In classical masked autoencoder approaches, hidden pixel patches serve as the training signal: the model reconstructs masked image regions and forms a general understanding of visual content. LingBot-Vision switches the signal — object boundaries themselves become the target information.

The idea comes from a simple observation: contours between objects carry dense spatial information — where an object ends, how the scene geometry is arranged, what the shape of surfaces is. The model learns to recognize and reconstruct these contours. For robotic systems, this is a critically important type of knowledge: a robot needs not just to recognize that there is a table in front of it, but to precisely understand where it begins and ends in space.

The method is fully self-supervised: manual annotation is not required. The model extracts the training signal directly from the structure of images — data preparation is fundamentally cheaper compared to supervised approaches.

Why a 1B model outperforms larger analogues

Robbyant's key result: the 1B-backbone LingBot-Vision on dense spatial perception tasks matches or exceeds larger models. Notably, this is about ViT architecture: Vision Transformer backbones typically benefit from scaling — therefore, high results at 1 billion parameters speak primarily to the efficiency of the training signal itself.

Key model characteristics:

• Number of parameters: 1 billion (1B ViT backbone)
• Training method: masked boundary modeling without manual annotation
• Target task: dense spatial perception
• Application in ecosystem: initialization of LingBot-Depth 2.0
• Public release date: July 7, 2026

Dense spatial perception covers depth estimation, surface geometry understanding, and precise object detection in three-dimensional space — boundary information provides maximum benefit precisely there. LingBot-Vision also serves as the weight base for LingBot-Depth 2.0, a depth perception system from the same Robbyant ecosystem, indicating a systematic approach: Ant Group consistently builds an interconnected stack of vision components for robotic applications.

Open code for niche development

Ant Group is a fintech giant, the parent company of Alipay, one of China's largest technology players. Robbyant is its division focused on robotic systems and AI for the physical world.

The public release of LingBot-Vision fits into a notable strategy of major Chinese AI labs: publishing competitive models in open access, establishing standards in target niches. Publishing code and weights lowers the entry barrier for teams working on robotic perception, industrial computer vision, and autonomous systems. The masked boundary modeling approach is meanwhile open for adaptation: researchers can experiment with other types of structural signals on top of the same architectural base.

What this means

LingBot-Vision demonstrates: in spatial perception tasks, the correct training signal matters more than model scale. For engineers and researchers working on 3D perception and robotics, this is a ready-made open-source foundation without costly data annotation. Masked boundary modeling as a method is promising in itself: it opens opportunities for experimentation with types of structural information extracted from images without annotation.