Robbyant, part of Ant Group, has introduced LingBot-Depth 2.0, a spatial perception model for robotics, alongside LingBot-Vision, a visual foundation model designed to support robotic perception and downstream computer vision tasks. LingBot-Depth 2.0 builds on the company’s earlier LingBot-Depth model, which used Masked Depth Modeling (MDM) to address depth sensing challenges involving transparent and reflective surfaces. According to Robbyant, the new model was trained on 150 million samples and achieved the highest ranking in 12 of 16 depth completion benchmarks.
In indoor scenarios characterized by substantial depth loss, the company said the model reduced root mean square error (RMSE) from 0.132 to 0.062 compared with its predecessor. Robbyant also stated that the model improves depth reconstruction for glass, mirrors and other transparent objects, which have traditionally presented challenges for depth cameras.
The company said LingBot-Vision underpins these improvements by using boundary structure as its pre-training objective. Robbyant described the model as providing sub-pixel boundary localization and spatial structure understanding to improve visual representations for robotic perception. The model was trained on a dataset of 160 million images, which the company said is smaller than the training corpus used for DINOv3 while supporting stable object boundary detection and continuous boundary tracking in video.
In addition to supporting LingBot-Depth 2.0, Robbyant said LingBot-Vision is intended for use as a general-purpose visual foundation model across multiple downstream applications.
Robbyant also outlined a collaboration with Orbbec, a supplier of robotics and artificial intelligence vision systems. The company said LingBot-Depth 2.0 has been certified by Orbbec’s Depth Vision Laboratory and evaluated using chip-level depth data from the Gemini 330 series stereo 3D cameras. According to Robbyant, testing showed improvements in edge definition, object contour reconstruction, recognition of small objects, long-range depth estimation, and performance under varying lighting and material conditions.
The companies also plan to integrate a customized version of LingBot-Depth into Orbbec’s RGB-D EGO device, part of the Robot-Free Data Collection Hardware Platform, with the aim of improving depth completion and spatial data collection for embodied AI training.
Robbyant and Orbbec said they will introduce two products incorporating LingBot-Depth 2.0: a software development kit for robotics systems using the Gemini 330 camera series and an integrated camera combining 3D imaging and spatial perception capabilities, which is expected to be released by the end of the year.
Robbyant also said it has open-sourced the model weights for LingBot-Vision as part of its effort to support broader development of robotic vision technologies.
