Chinese AI company Robbyant has launched LingBot-Depth 2.0, a next-generation spatial perception model designed to enhance robotic navigation in complex environments. This model builds on the previous LingBot-Depth, utilizing the Masked Depth Modeling technique and trained on 150 million samples, achieving top results in 12 out of 16 depth completion benchmarks. Notably, it reduces depth error by over 50%, improving accuracy in challenging indoor settings.
The significance of LingBot-Depth 2.0 lies in its ability to accurately perceive transparent and reflective surfaces, areas where traditional depth cameras often struggle. The model's advancements are attributed to LingBot-Vision, a visual foundation model that employs a unique “boundary structure” pre-training objective, enabling sub-pixel-level boundary localization. Despite being trained on a smaller dataset of 160 million images, it demonstrates robust performance across various robotic vision applications, enhancing object boundary detection and tracking.
Looking ahead, Robbyant's collaboration with Orbbec aims to integrate LingBot-Depth 2.0 into new hardware solutions for robotics data collection. The RGB-D EGO device, part of Orbbec’s Robot-Free Data Collection Hardware Platform, will feature a customized version of the model. Future updates are expected to further enhance depth completion and spatial structure understanding, providing a solid foundation for training embodied AI systems in real-world scenarios. No further timeline was disclosed at the time of publication.
Editor's Note
The launch of LingBot-Depth 2.0 highlights a significant trend in the robotics industry towards enhancing spatial perception capabilities. As companies like Robbyant innovate in depth sensing technologies, we may see a shift in how robots interact with complex environments. This could lead to increased adoption of advanced perception models across various sectors, including logistics and autonomous vehicles.
Leave a comment