Ant Group's Lingbo Tech Releases LingBot-Depth 2.0 and Open-Sources LingBot-Vision Foundation Model

On July 7, Ant Group's embodied AI company Lingbo Technology officially launched LingBot-Depth 2.0, its next-generation spatial perception model that serves as the visual foundation for robots navigating complex physical environments. The model addresses long-standing challenges in perceiving transparent objects, reflective surfaces, and areas with large depth gaps. Compared to version 1.0, training data expanded from 3 million to 150 million samples, driving substantial performance improvements across the board.

In the most difficult scenario — indoor environments with large depth-missing regions — depth error was cut in half (RMSE dropping from 0.132 to 0.062). The model performed especially well on glass, mirrors, and transparent objects, where traditional depth cameras typically fail. Across all 16 depth completion benchmark evaluations, LingBot-Depth 2.0 achieved 12 first-place results. It has also passed professional certification from Orbbec's depth vision lab.

Alongside the depth model, Lingbo released LingBot-Vision, a visual foundation model that claims a training paradigm breakthrough: it is the first industry model to use "boundary structure" as a pre-training objective. LingBot-Vision achieves sub-pixel boundary localization and spatial structure understanding, using only 160 million pre-training images — an order of magnitude fewer than DINOv3 — while delivering superior depth estimation accuracy.

LingBot-Vision has been open-sourced in four versions: ViT-G, ViT-L, ViT-B, and ViT-S. Beyond powering LingBot-Depth 2.0's training pipeline, the model supports general-purpose visual tasks across multiple robot vision scenarios, making it a versatile tool for the embodied AI ecosystem.

On the commercialization front, Lingbo has entered into a deep partnership with Orbbec. Orbbec will release an SDK integrating the latest LingBot-Depth model capabilities for robot clients to use on edge devices. By year-end, the companies plan to launch an integrated camera product combining 3D hardware with LingBot-Depth's commercial edition.

The release and open-sourcing of both models represents a significant step in Ant Group's embodied AI strategy. By sharing model weights and technical reports openly, Lingbo aims to co-build the robotic vision foundation with the broader industry, accelerating the path from lab research to real-world deployment.