Alibaba’s Amap Introduces Full Stack Upgrade to ABot Embodied AI System

Amap, the location-based services platform of Alibaba Group, announced in a press release a full stack upgrade to its ABot embodied AI system. The new release includes ABot-N1, ABot-M0.5, ABot-ER, ABot-AgentOS, and ABot-C0, aimed at improving navigation, manipulation, reasoning, memory, and motion control capabilities in robots.

ABot-N1 is a general navigation foundation model that enables robots to move through open environments using conventional human navigation maps. It employs a dual system design with a slower module for reasoning and a faster one for control, reaching a 92.9% outdoor navigation success rate according to Amap.

ABot-M0.5 serves as a general manipulation model that manages robot movement and object interaction. It introduces a training process called dream self-healing to reduce operational errors, surpassing previous results on RoboCasa-365 tests by more than 20% on complex tasks.

The embodied-agent layer consists of ABot-ER and ABot-AgentOS. ABot-ER handles perception-based reasoning and achieved leading results on three benchmarks including Embodied Arena 2D-EQA. ABot-AgentOS converts planning into actions across various robot forms such as humanoid, quadruped, and wheeled designs, and incorporates a lifelong multimodal memory that learns from previous tasks.

The ABot-C0 component translates system-level decisions into physical actions for quadruped robots, supporting coordination among different robot types. Together, the upgraded components allow robots to learn continuously through interaction and adaptation in physical environments. Research papers for the models have been published on arXiv.