Insider Brief
- Robbyant, Ant Group’s physical AI company, launched LingBot-Depth 2.0 and LingBot-Vision, two robot-vision models aimed at improving robotic depth perception and spatial understanding.
- Robbyant said LingBot-Depth 2.0 was trained on 150 million samples, achieved top rankings in 12 of 16 depth-completion benchmarks and reduced depth error in difficult indoor scenes compared with its predecessor.
- The company said LingBot-Depth 2.0 has been certified by Orbbec’s Depth Vision Laboratory, and the companies are working on commercial products including an SDK for Gemini 330 series cameras and an integrated camera planned for release by year-end.
Robbyant, Ant Group’s physical AI company, launched two robot-vision models designed to improve how robots understand depth and physical space.
LingBot-Depth 2.0, a spatial-perception model, and LingBot-Vision, a visual foundation model are geared at helping robots better interpret and navigate real-world environments, according to the company.
LingBot-Depth
LingBot-Depth 2.0 builds on an earlier Robbyant model that used masked depth modeling to address problems with transparent and reflective surfaces. The company said the new model was trained on 150 million samples and achieved top rankings in 12 of 16 depth-completion benchmarks.
The model showed its strongest improvements in difficult indoor scenes with large amounts of missing depth information. In those cases, the company said LingBot-Depth 2.0 cut depth error compared with the previous model, reducing root mean square error to 0.062 from 0.132.
The company indicated the model also improved performance in cases that often challenge conventional depth cameras, including glass, mirrors and transparent objects.
LingBot-Vision
LingBot-Depth 2.0 is supported by LingBot-Vision, a visual foundation model that uses boundary structure as a pre-training objective. That approach helps the system identify object boundaries and understand spatial structure with high precision, according to Robbyant.
LingBot-Vision was trained on 160 million images and the company said the model can also support downstream vision tasks beyond LingBot-Depth 2.0 and can track object boundaries in video.
Robbyant also noted that LingBot-Depth 2.0 has been certified by Orbbec’s Depth Vision Laboratory. In real-world testing using chip-level depth data from Orbbec’s Gemini 330 series stereo 3D cameras, Robbyant reported that the model improved edge clarity, object contours, small-object recognition, long-range depth estimation and performance under complex lighting and material conditions.
Orbbec’s RGB-D EGO device, part of its Robot-Free Data Collection Hardware Platform, will integrate a customized LingBot-Depth model for data collection, according to Robbyant. The companies also plan to integrate a commercial version aimed at improving missing-depth completion, object edges and spatial structures for embodied AI training data, Robbyant said.
Orbbec Collaboration
Robbyant said it is working with Orbbec on two commercial products tied to LingBot-Depth 2.0. One is an SDK for edge deployment by robotics customers using Gemini 330 series cameras. The other is an integrated camera, planned for release by year-end, that combines a 3D camera with spatial-perception software.
The Chinese embodied AI company said it has also open-sourced the model weights for LingBot-Vision on GitHub and Hugging Face.
Image credit: Robbyant
