高德发布全栈具身导航基座模型ABot-N0，五大核心任务全面SOTA

Core Insights - Gaode has launched two foundational models, ABot-M0 and ABot-N0, achieving state-of-the-art (SOTA) performance in both embodied navigation and operation tasks, marking a significant milestone in the industry [1] Group 1: ABot-M0 Overview - ABot-M0 is the world's first unified architecture for robotic operation foundational models, integrating over 6 million real operation trajectories to create the largest general-purpose robotic dataset [1] - The model utilizes action manifold learning algorithms for cross-platform action prediction, achieving a task success rate of 80.5% in the Libero-Plus benchmark test, which is nearly a 30% improvement over the industry solution pi0 [1] Group 2: ABot-N0 Overview - ABot-N0 is the first full-stack navigation foundational model that integrates five major navigation tasks, including point navigation, target navigation, instruction following, point of interest navigation, and person following [1] - The model is trained on approximately 8,000 high-fidelity 3D scenes and nearly 17 million expert examples, employing a hierarchical "brain-action" architecture [1] - In benchmark tests such as CityWalker and SocNav, ABot-N0 has set new world records, with a 40.5% success rate improvement in closed-loop simulation for SocNav and an 8.8% success rate increase in the HM3D-OVON evaluation [1][2]