en.Wedoany.com Reported - Amap released a full-stack upgrade of its ABot embodied AI system in Beijing, China on July 22, 2026. Amap, an Alibaba-affiliated location-based service platform, unveiled five components—ABot-N1, ABot-M0.5, ABot-ER, ABot-AgentOS, and ABot-C0—achieving state-of-the-art (SOTA) results across 17 widely used benchmarks. The system aims to address a key bottleneck in embodied intelligence: enabling robots to perceive, reason, and act within physical environments.
The ABot architecture is built as a full-stack embodied intelligence technology system, connecting world models, foundation models, and an embodied agent framework, allowing data from simulated training, physical interaction, and memory scheduling to feed back into the system and improve performance over time. ABot-N1 is a general-purpose navigation foundation model with a dual-system architecture: a slower module handles long-range reasoning, while a faster module manages real-time control; it introduces a chain-of-thought process at the pixel level, combining visual input with language-based logic to make decision-making more traceable. The model can achieve city-scale autonomous navigation using only navigation maps, and according to Amap, it leads comparable systems in point-goal navigation, object-goal navigation, instruction following, point-of-interest navigation, and person following, with an outdoor navigation success rate of 92.9%.
ABot-M0.5 is a general-purpose manipulation foundation model that separates movement and manipulation into two action streams, using frame-level implicit actions to align vision, hands, and movement. It employs a "dream self-healing" training method, enabling it to continue generating actions from noisy visual predictions, reducing failures caused by minor deviations. In the RoboCasa-365 test, ABot-M0.5 outperformed the previous state-of-the-art by 20.4% on complex tasks and 10.6% on basic tasks.
ABot-ER supports decision-making from perception to action, helping robots reason about relationships between objects, spaces, and tasks. As of its release, ABot-ER achieved SOTA results on three benchmarks and ranked first on the Embodied Arena 2D-EQA benchmark released by multiple research institutions. ABot-AgentOS translates plans into executable actions, supporting humanoid, quadruped, and wheeled robots, and provides multimodal lifelong memory that can be optimized based on failed task trajectories. Together, they convert task experience into long-term memory and decision-making references while protecting local private memory security.
ABot-C0 is a motion control component that translates ABot system decisions into physical actions, building a unified behavior foundation for quadruped robots and supporting collaboration across heterogeneous robot types. The upgraded ABot system is designed to help robots continuously learn and adapt in real-world open environments. Amap has published research papers on ABot-N1, ABot-M0.5, ABot-AgentOS, and ABot-C0 on arXiv.
