According to the official website of Obi China Light, on July 7th, Ant Group's embodied intelligence company Ant Lingbo released a new generation of spatial perception model LingBot Depth 2.0. This model, combined with Obi China Light cameras, significantly improves spatial perception capabilities in complex real-world environments and has obtained certification from the Obi China Light Deep Vision Laboratory. At the same time, Obi Zhongguang has officially launched a non ontology data acquisition hardware platform, and its core product EGO RGB-D is adapted to this model, jointly building a high-quality data acquisition base for embodied intelligent training.

The ontology free data acquisition hardware platform launched by Obi Zhongguang this time integrates product forms such as EGO, UMI, WristCam, etc., covering typical scenarios such as first angle observation, wrist near-field observation, and hand object interaction detail acquisition, providing standardized products and CM, JDM services for embodied intelligent model companies, robot ontology companies, etc.

Among them, EGO RGB-D, which is designed for first person interactive data collection, is a "chip level depth direct output+model level depth enhancement" solution jointly created by both parties. On the hardware side, the product integrates the Gemini 330 series binocular 3D camera equipped with the self-developed depth engine chip MX6800 by Obi Zhongguang. It can synchronously capture RGB images and high-precision depth data, and has the ability to output depth directly with low latency and stable output. On the model side, EGO RGB-D is based on the original depth data and will be adapted to LingBot Depth 2.0, which is specifically optimized for data collection scenarios using Ant Lingbo. This will further fill in depth gaps, optimize object edges and spatial structure details, and improve the quality and availability of depth maps in complex scenes such as reflective, transparent objects, and occluded edges. Both parties will also adapt to higher-level commercial versions of the model in the future.

In a wider range of robot vision applications, both parties are promoting the deep integration of high-precision 3D vision hardware and advanced AI algorithm models on the end side. The LingBot Depth 2.0 released this time is based on 150 million scale data training, and has been comprehensively upgraded in edge clarity, small object recognition, long-range depth estimation, and robustness to complex scenes. It achieved 12 first places in 16 evaluations of the depth completion benchmark. In indoor large-scale depth missing scenes, the depth error (RMSE) decreased from 0.132 in the previous generation to 0.062, and it performed outstandingly in scenes where traditional cameras such as glass and mirrors are prone to failure.

The cooperation between the two parties can be traced back to the LingBot Depth 1.0 stage. Prior to this, the LingBot depth enhancement engine was developed specifically for Gemini 330 series cameras to address difficulties such as transparent objects, weak textures, and strong reflective materials. Next, Obi Zhongguang will launch an SDK product that integrates LingBot Depth's latest model capabilities, and plans to launch an integrated camera product that integrates LingBot Depth's commercial version as early as the end of the year, achieving the integrated delivery of "3D camera+spatial perception capability".

Original text: Obi Zhongguang releases data collection platform, collaborates with Ant Lingbo's latest model to accelerate robot scene landing (source: Obi Zhongguang official website)