Jin Rong Jie·2026-02-12 07:31Core Insights - Alibaba's Gaode has launched two foundational models, ABot-M0 and ABot-N0, enhancing the capabilities of embodied robots in operation and navigation, achieving state-of-the-art (SOTA) performance in both areas [1][9] Group 1: ABot-M0 Overview - ABot-M0 is the world's first unified architecture for a robotic foundational model, providing a "general brain" for robots [2] - The model addresses challenges in robotic technology by restructuring data, algorithms, and spatial perception to improve generalization across diverse robot forms and tasks [5] - It integrates over 6 million real operation trajectories to create the largest general robotic dataset, enabling cross-platform data fusion and pre-training based on public data [5] - The model introduces Action Manifold Learning (AML) to predict physically feasible action sequences, enhancing stability and decoding efficiency [5] - In benchmark tests like Libero and Libero-Plus, ABot-M0 achieved an average task success rate of 80.5%, nearly 30% higher than the previous leading solution [5][7] Group 2: ABot-N0 Overview - ABot-N0 is the first all-stack navigation foundational model that integrates five major navigation tasks: Point-Goal, Object-Goal, Instruction-Following, POI-Goal, and Person-Following [9][10] - This model overcomes fragmentation in existing navigation research, allowing for better task generalization and execution of complex commands [9] - ABot-N0 employs a hierarchical "brain-action" design, utilizing cognitive training and reinforcement learning to align navigation decisions with human preferences [12] - It leverages a vast navigation data engine with approximately 8,000 high-fidelity 3D scenes and nearly 17 million expert examples, enhancing robustness in real environments [12] - The model set new world records in seven authoritative benchmark tests, with a 40.5% increase in success rate in SocNav simulations and an 8.8% increase in HM3D-OVON evaluations [12][14]