Space intelligence has made significant open-source progress. On January 27, Lingbo Technology, a embodied intelligence company under Ant Group, announced the open-source release of the high-precision spatial perception model LingBot-Depth.

This model is based on chip-level raw data provided by the Orobote Gemini330 series stereo 3D camera, focusing on improving environmental depth perception and three-dimensional spatial understanding capabilities. It aims to endow intelligent terminals such as robots and autonomous vehicles with more accurate and reliable three-dimensional vision, achieving an important breakthrough in the industry's key challenge of "seeing clearly" in a three-dimensional world. This is also the first major achievement in the field of embodied intelligence technology base announced by Ant Lingbo Technology half a year after its debut at the 2025 Outer Bank Conference.

In authoritative benchmark evaluations such as NYUv2 and ETH3D, LingBot-Depth demonstrates a generation-level advantage: compared to mainstream industry solutions like PromptDA and PriorDA, its relative error (REL) in indoor scenarios is reduced by more than 70%, and the RMSE error in challenging sparse SfM tasks is reduced by about 47%, establishing a new industry precision standard.

QQ20260127-122552.png

(Figure caption: In the most challenging sparse depth completion task, LingBot-Depth performs better than various existing mainstream models. The lower the number in the figure, the better the performance.)

In home and industrial environments, transparent and reflective objects such as glassware, mirrors, and stainless steel equipment are very common but pose challenges for machine space perception. Traditional depth cameras are limited by optical physical characteristics and often fail to receive effective echoes when facing transparent or highly reflective materials, leading to data loss or noise in depth maps.

To address this common industry challenge, Ant Lingbo Technology has developed the "Masked Depth Modeling" (MDM) technology and used the Orobote Gemini330 series stereo 3D camera for RGB-Depth data collection and validation. When depth data is missing or abnormal, the LingBot-Depth model can infer and complete the missing areas by integrating texture, contour, and contextual information from color images (RGB), outputting a complete, dense, and edge-clearer 3D depth map. Notably, the LingBot-Depth model has passed professional certification by Orobote's depth vision laboratory, achieving industry-leading levels in accuracy, stability, and adaptability in complex scenarios.

In experiments, the Orobote Gemini330 series, when applying LingBot-Depth, still outputs smooth and complete depth maps even in challenging optical scenarios such as transparent glass, highly reflective mirrors, strong backlighting, and complex curved surfaces. The object outlines are very sharp, and its performance significantly exceeds that of Stereolabs' ZED Stereo Depth 3D camera, a leading 3D vision company in the industry. This means that without replacing sensor hardware, LingBot-Depth can significantly improve the processing effect of consumer-grade depth cameras on difficult objects.

5bb53c659f33ffc0aefe4508f26d8607.png

(Figure caption: [Top image] After applying LingBot-Depth, the Orobote Gemini330 series shows significantly improved completeness and edge clarity of depth maps in transparent and reflective scenes; [Bottom image] Its performance surpasses that of the leading ZED depth camera in the industry)

The excellent performance of LingBot-Depth comes from massive real-world scene data. Lingbo Technology has collected approximately 10 million original samples and distilled 2 million high-value depth pairs for training, supporting the model's generalization capability under extreme conditions. This core data asset (including 2 million real-world depth data and 1 million simulation data) will be open-sourced soon, helping the community overcome complex scene perception challenges more quickly.

It is understood that Ant Lingbo Technology has reached a strategic cooperation intention with Orobote. Orobote plans to launch a new generation of depth cameras based on LingBot-Depth's capabilities.

This week, Ant Lingbo will gradually open-source multiple models in the direction of embodied intelligence.