Ant Group's Robbyant open-sources LingBot-Vision, a boundary-centric vision model for embodied AI
marktechpost.com

Ant Group's Robbyant open-sources LingBot-Vision, a boundary-centric vision model for embodied AI

Tech News
2 min read

Published by AINave Editorial • Reviewed by Ramit

TL;DRAnt Group's Robbyant open-sources LingBot-Vision, a family of self-supervised Vision Transformers with a boundary-centric pretraining objective. The flagship 1.1B-parameter model matches or beats 7B DINOv3 on dense spatial tasks like depth estimation and video object segmentation, trained on an order of magnitude less data.

Most vision foundation models learn what is in an image while discarding the fine-grained spatial structure that robots and other physically embodied systems depend on. LingBot-Vision inverts that priority: it treats object boundaries as a native pretraining signal rather than a downstream output. The result is a 1.1B-parameter backbone that matches or surpasses models up to 7x larger on dense spatial tasks, including the 7B DINOv3.

What happened

Robbyant, the embodied-AI company within Ant Group, open-sourced LingBot-Vision under Apache-2.0 on Hugging Face in four Vision Transformer sizes: ViT-giant (1.1B), ViT-large (300M), ViT-base (86M), and ViT-small. The release includes a technical report and inference code.

The flagship ViT-g/16 uses a novel masked boundary modeling objective trained on about 161M images curated from a 2B web pool. Importantly, the training corpus is an order of magnitude smaller than DINOv3's LVD-1689M, and the model consumes less than a third of DINOv3's training samples. No human labels, no external edge detectors, and no pretrained backbone to bootstrap from.

LingBot-Depth 2.0 extends the depth-completion system using LingBot-Vision initializations. The same masked-depth-modeling recipe from version 1.0 uses LingBot-Vision encoder initialization (ViT-L and ViT

Sources

Latest Tech News