ABot-AgentOS Introduces General Robotic Agent Operating System with Memory

While recent advancements in Vision-Language Models (VLMs) and Vision-Language-Action (VLA) systems have improved robotic perception and action, complex, long-duration embodied agents still lack a comprehensive runtime layer for advanced reasoning, memory management, and tool integration. To address this, researchers have developed ABot-AgentOS, a novel robotic Agent Operating System. ABot-AgentOS functions as a deliberative layer above low-level controllers, enabling capabilities such as scene-conditioned planning, context-isolated skill execution, multi-stage verification, and seamless edge-cloud collaboration. A core component is its Universal Multi-modal Graph Memory, a persistent system that converts various inputs—dialogue, visual observations, spatial context, temporal relations, and task traces—into structured nodes and edges, providing an auditable and continually evolving knowledge base. The system also incorporates a failure-driven self-evolution loop. This mechanism diagnoses memory failures and promotes "evo-assets" (improved runtime components) only to future evaluation splits, preventing data leakage while fostering continuous learning and improvement. Evaluated on a new benchmark called EmbodiedWorldBench, ABot-AgentOS demonstrated improved task success and goal completion compared to single-controller baselines, highlighting the benefits of a general agent OS layer for complex robotic tasks.