- LingBot-Depth 2.0 improves depth sensing in difficult environments such as glass and reflective surfaces
- The model upgrade comes with a vision foundation model and a hardware partnership to accelerate commercialization
Robbyant (蚂蚁灵波科技), Ant Group’s embodied intelligence subsidiary, has released LingBot-Depth 2.0, an upgraded spatial perception model, alongside its vision foundation model LingBot-Vision, aiming to help robots better understand and interact with the physical world.
Robots have long struggled with visual perception in challenging environments involving transparent glass, reflective surfaces, small objects and complex lighting conditions.
In scenarios such as identifying a glass champagne tower, conventional depth cameras often produce fragmented or incomplete depth maps, making it difficult for robots to accurately locate objects, grasp them or move reliably.
LingBot-Depth 2.0 was developed to address these limitations.
Ranking first in 12 benchmarks
Compared with its previous version, the model expands its training dataset from 3 million to 150 million samples and ranked first in 12 out of 16 depth completion benchmarks, according to Robbyant.
In indoor environments with large areas of missing depth information, one of the most challenging perception scenarios, the model reduced depth errors by more than half, with Root Mean Square Error (RMSE) improving from 0.132 to 0.062.
It also delivered stronger performance in scenes involving glass, mirrors and transparent objects, where traditional depth sensing systems often fail.
LingBot-Vision
Alongside LingBot-Depth 2.0, Robbyant also launched LingBot-Vision, a vision foundation model designed to improve robots’ understanding of spatial structures and object boundaries.
The company said LingBot-Vision is the first vision model to use boundary structures as a pre-training objective, enabling subpixel-level boundary localization and stronger spatial understanding.
With less than one-third of the training data used by Meta AI’s DINOv3 self-supervised vision model, LingBot-Vision has surpassed it across multiple standard vision benchmarks, according to Robbyant.
The model has been open-sourced in four versions — ViT-G, ViT-L, ViT-B and ViT-S — for developers and researchers.
Commercialization progress
For commercialization, Robbyant has partnered with 3D vision hardware maker Orbbec (奥比中光). The company plans to integrate LingBot-Depth capabilities into SDK products and launch commercial integrated camera solutions combining 3D sensing hardware with spatial perception models by the end of the year.
The partnership marks Robbyant’s move from open-source spatial intelligence models toward integrated hardware and software solutions, as the company seeks to build a technology pipeline enabling robots to better perceive and operate in real-world environments.
