Amap, Alibaba’s location-based services platform, has recently unveiled a comprehensive update to its ABot embodied AI system, launching ABot-N1, ABot-M0.5, ABot-ER, ABot-AgentOS, and ABot-C0. These advancements aim to tackle crucial challenges in embodied intelligence, focusing on enabling robots to sense, reason, and act in physical environments. The updated ABot architecture establishes a groundbreaking full-stack embodied-intelligence technology system, linking world models, foundation models, and an embodied-agent framework.
One of the key components of the upgrade is ABot-N1, a foundational model for general navigation intended to assist robots in traversing open environments. Unlike traditional navigation maps designed for humans, ABot-N1 integrates a dual-system architecture that combines long-range reasoning with real-time control to enhance spatial perception. By incorporating a pixel-level decision-making process that merges visual inputs with language-based logic, ABot-N1 can autonomously navigate city-scale environments with impressive success rates.
Another significant addition is ABot-M0.5, a manipulation foundation model that streamlines robot movement and object interaction in real-world scenarios. By separating locomotion and manipulation into distinct action streams and implementing frame-level implicit actions, ABot-M0.5 enhances synchronization between vision, hands, and movement. Through a “dream self-healing” training approach, the model optimizes action generation from visual predictions, minimizing failures and improving task performance significantly.
ABot-ER and ABot-AgentOS constitute the embodied-agent layer of the system, facilitating decision-making and planning for robots. ABot-ER enhances robots’ ability to reason about object relationships and spatial tasks beyond visual recognition, achieving state-of-the-art results across various benchmarks. On the other hand, ABot-AgentOS translates planning into actionable steps, supporting different robot forms and enabling multimodal lifelong memory for efficient task execution and skill transfer.
Furthermore, ABot-C0 serves as the motion-control component, translating decisions from the ABot system into physical actions and enabling collaboration among diverse robot types. The upgraded ABot system emphasizes continual learning and adaptation in real-world environments, with a focus on sharing knowledge and experiences across different robot models.
Research papers detailing the advancements of ABot-N1, ABot-M0.5, ABot-AgentOS, and ABot-C0 are available on arXiv, showcasing the innovative strides made by Amap to revolutionize embodied intelligence in robotics.
About the Author
