Ant Robbyant Open-Sources LingBot Vision Models, Topping Depth Benchmarks

Ant Group's embodied AI unit, Robbyant, officially released and open-sourced LingBot-Vision, a visual foundation model family designed for robots, alongside the depth perception model LingBot-Depth 2.0. Released under the Apache-2.0 license, LingBot-Vision offers four backbone sizes ranging from 21M to 1.1B parameters (ViT-S to ViT-g). The models demonstrate outstanding performance in dense spatial perception tasks, with LingBot-Depth 2.0 securing 12 first-place rankings on industry benchmarks. This release marks a significant breakthrough in foundational visual perception for embodied AI.

Technical Innovation and Core Advantages

LingBot-Vision is a self-supervised vision backbone belonging to the DINO lineage. Its core innovation is "Masked Boundary Modeling." According to interpretations by @StillThese3747 and @Illustrious-Data1712, this mechanism uses a teacher model to online-predict dense boundary fields, forcing tokens containing boundaries into the student network's mask. This compels the student to reconstruct areas that cannot be inferred from context, requiring no manual annotation or external edge detectors. For training, the model utilized 161 million images filtered from 2 billion original pictures.

Performance and Data Scaling

Regarding performance, @thetripathi58 noted that the 1B LingBot-Vision achieves a lower depth estimation RMSE on the NYU-Depth v2 benchmark than Google SigLIP 2 and NVIDIA AM-RADIO. It improves overall depth estimation by roughly 40% over Google's model and adds 10.8 mIoU on the ADE20K segmentation benchmark. Furthermore, the model exhibits excellent scaling characteristics: @AdinaYakup mentioned its 1.1B parameter model outperforms the 7B DINOv3, while @thetripathi58 added that expanding depth training data to 150 million samples continuously improves initialization. Even when distilled into a 0.3B student model, it remains highly competitive on edge devices.

Tackling Complex Scenarios and Open-Sourcing Strategy

For real-world robotic applications, LingBot-Depth 2.0 focuses on solving broken depth maps and hollow issues caused by transparent/reflective materials, small targets, and low-light environments. @Ok-Line2658 analyzed that its approach uses invalid depth signals generated by sensors in specular highlights or texture-less areas directly as masking signals for training. Regarding the open-source strategy, @thetripathi58 confirmed that LingBot-Vision's weights and code are open, but the downstream LingBot-Depth 2.0 remains closed-source.

2026-07-07 ~ 2026-07-08 · 18 related posts

1 near-duplicate retellings: AdinaYakup