让机器人“看清”三维世界 蚂蚁灵波开源空间感知模型

Core Insights - Ant Group's Lingbo Technology has made significant advancements in spatial intelligence by open-sourcing the high-precision spatial perception model LingBot-Depth, aimed at enhancing depth perception and 3D spatial understanding for robots and autonomous vehicles [1] Group 1: Model Capabilities - LingBot-Depth utilizes raw data from the Orbbec Gemini 330 series dual-camera 3D system, achieving over a 70% reduction in relative error (REL) in indoor scenes compared to mainstream models like PromptDA and PriorDA [1] - The model demonstrates a 47% reduction in RMSE error in challenging sparse Structure from Motion (SfM) tasks, showcasing its generational advantage in performance [1] Group 2: Addressing Industry Challenges - Traditional depth cameras struggle with transparent and reflective objects, leading to data loss or noise in depth maps. Lingbo Technology has developed "Masked Depth Modeling" (MDM) to address this issue [3] - LingBot-Depth can infer and complete missing depth data by integrating texture, contours, and contextual information from RGB images, resulting in clearer and more complete 3D depth maps [3] Group 3: Performance Validation - In tests, the Gemini 330 series, when paired with LingBot-Depth, produced smooth and complete depth maps even in challenging optical scenarios, outperforming Stereolabs' ZED Stereo Depth camera [4] - This indicates that LingBot-Depth can significantly enhance the performance of consumer-grade depth cameras without requiring hardware changes [4] Group 4: Data and Collaboration - The model's effectiveness is supported by a vast dataset, with approximately 10 million raw samples and 2 million high-value depth pairs used for training, enhancing its generalization capabilities in extreme environments [6] - Ant Group's Lingbo Technology has established a strategic partnership with Orbbec to develop a new generation of depth cameras based on LingBot-Depth's capabilities, with plans to open-source multiple models in the field of embodied intelligence [6]