Ant Group's Robbyant Open-Sources LingBot-Vision Model

Robbyant, an embodied-AI company within Ant Group, has open-sourced LingBot-Vision, a family of self-supervised Vision Transformers specifically engineered for dense spatial perception. The model weights are available under the Apache-2.0 license on Hugging Face in four sizes: ViT-giant, ViT-large, ViT-base, and ViT-small. This release also includes a technical report and inference code.

Unlike many existing vision foundation models that prioritize semantic invariance and discard fine-grained spatial details like object boundaries and depth discontinuities, LingBot-Vision inverts this focus. It treats boundaries as a primary pretraining signal. This approach allows its 1 billion parameter backbone to match or exceed the performance of models up to 7 times larger on dense spatial tasks, including the 7B DINOv3 model.

The flagship ViT-g/16 model, with approximately 1.1 billion parameters, was trained using a novel objective called masked boundary modeling. This training utilized a curated corpus of around 161 million images, selected from a larger pool of 2 billion web images, without requiring human labels, external edge detectors, or a pretrained backbone for bootstrapping. The training process is also noted for its efficiency, using a corpus an order of magnitude smaller than DINOv3's LVD-1689M and consuming less than a third of DINOv3's training samples.

LingBot-Vision's encoder generates dense patch-token features designed for frozen readouts. For applications with more constrained budgets, the flagship model has been distilled into smaller student models: ViT-L (300 million parameters), ViT-B (86 million parameters), and ViT-S. These distilled versions are designed to lead in dense prediction tasks within their respective size classes, making advanced spatial perception capabilities more accessible.