Robbyant, an embodied artificial intelligence company under Ant Group, has released LingBot-Depth, an open-source spatial perception model designed to enhance robots’ depth sensing and 3D environmental understanding. The company developed the model with technical support from Orbbec, a provider of robotics and AI vision systems, which plans to integrate the technology into its next-generation depth cameras for embodied intelligence applications.
According to Robbyant, LingBot-Depth delivers improved accuracy over existing models on benchmark datasets such as NYUv2 and ETH3D. The company reports that the model reduces relative error in indoor scenes by more than 70 percent and lowers root mean square error by about 47 percent in sparse depth completion tasks compared with models including PromptDA and PriorDA.
The model is based on Robbyant’s Masked Depth Modeling (MDM) framework, which reconstructs missing or corrupted depth information from RGB image features when sensors fail to capture reliable data on transparent or reflective surfaces. This approach enables the generation of denser and more accurate 3D depth maps without requiring modifications to existing hardware.
LingBot-Depth was co-optimized and validated on Orbbec’s platforms using data from the company’s Gemini 330 stereo 3D cameras, powered by the MX6800 depth engine chip. The collaboration leveraged chip-level raw depth data to improve depth map reconstruction and robot perception in complex optical environments.
Robbyant trained LingBot-Depth on approximately 10 million raw samples and a curated dataset of 2 million RGB-depth pairs collected under varied and extreme conditions. The company also intends to release this dataset publicly to support further research in spatial perception.
Robbyant Chief Executive Officer Zhu Xing said the initiative aims to advance embodied intelligence by improving accessibility to high-quality 3D vision tools. Len Zhong, Head of Product Management at Orbbec, described the collaboration as an example of integrating hardware depth sensing with perception algorithms to enhance robotic vision.
Robbyant stated that it plans to expand its spatial perception technologies to additional hardware partners to support deployment in diverse robotic environments such as homes, factories, and warehouses. The model and related resources are available on GitHub and Hugging Face.
