Jiayi Tian, Shiao Liu, Yuting Xu, Jia Lu, Zihao Guan, Honglin Han, Di Yang, Minqi Gu, Yifei Qian, Tianlin Zhang, Yanqing Zhu, Zeqian Ye, Menglin Yang, Fei Wang, Xu Hu, Xiuxian Li, Wei Zhang, Shihui Su, Yiyan Ji, Jingbo Wang, Ziteng Feng, Jiaheng Liu, Zhaoxiang Zhang, Xiaolong Wu, Mingyang Yin, Zedong Chu, Mu Xu
5 min
ABot-AgentOS is a robotic agent operating system designed to bridge the gap between high-level semantic reasoning and reliable physical execution. By sitting above low-level robot controllers, it provides a deliberative layer that manages task planning, skill execution, and multi-stage verification. The system is designed to be hardware-agnostic, allowing it to support various robot morphologies, from humanoids to mobile manipulators, by using a plugin-based skill integration paradigm.
The system employs a hierarchical architecture that separates the agent's responsibilities into three distinct roles: the main LLM, the Skill Runner, and the Verifier. The main LLM handles high-level scene-conditioned planning, while the Skill Runner manages procedural subtasks in an isolated context to prevent the main reasoning thread from becoming cluttered with low-level execution details. A critical component is the Verifier, which monitors execution to ensure progress and prevent premature termination, addressing common failure modes in long-horizon robotics like partial observability and actuation uncertainty.
ABot-AgentOS introduces a Universal Multi-modal Graph Memory that stores experiences as typed nodes and edges, capturing spatial, temporal, and semantic relations. This memory is persistent and source-grounded, allowing the agent to audit its own retrieval traces. Furthermore, the system implements a failure-driven self-evolution loop: when the agent encounters memory or execution failures, it diagnoses the cause and generates 'evo-assets'—gated runtime updates that improve future performance without leaking ground-truth data from current evaluation splits.
To evaluate these capabilities, the authors introduced EmbodiedWorldBench, a benchmark featuring 16 diverse scenes and over 200 tasks. Unlike traditional benchmarks that focus on isolated navigation or manipulation, this suite tests complex competencies such as NPC dialogue, dynamic event handling, and multi-step reasoning, providing a more rigorous standard for assessing the robustness of embodied agents.
Recent VLM and VLA systems have improved robotic perception and action prediction, yet long-horizon embodied agents still require a general runtime layer for reasoning, memory, tool use, verification, and cross-embodiment execution. We present ABot-AgentOS, a general robotic Agent Operating System that sits above low-level controllers and provides a deliberative agent layer for scene-conditioned planning, context-isolated skill execution, multi-stage verification, multi-modal memory, and edge-cloud collaboration. To evaluate such systems, we introduce EmbodiedWorldBench, an executable benchmark with 16 indoor, outdoor, and hybrid scenes, four difficulty levels, and over 200 tasks involving navigation, object search, NPC dialogue, dynamic events, and trace-grounded scoring. ABot-AgentOS further introduces Universal Multi-modal Graph Memory, a persistent source-grounded substrate that converts dialogue, visual observations, spatial context, temporal relations, and task traces into typed nodes and edges. A failure-driven self-evolution loop converts diagnosed memory failures into gated runtime evo-assets that are promoted only to later evaluation splits, preventing current-split ground-truth leakage while enabling continual improvement. On an initial EmbodiedWorldBench subset, ABot-AgentOS improves over a single-controller baseline in both task success and goal completion. Across memory benchmarks, ABot-AgentOS Static achieves 87.5 on LoCoMo, 59.9 on OpenEQA EM-EQA, 88.6 on Mem-Gallery, and 76.5 Acc@All on NExT-QA; self-evolution further improves LoCoMo to 88.7, OpenEQA to 60.4, and Mem-Gallery to 89.0. These results suggest that a general Agent OS layer can improve long-horizon embodied execution while providing persistent, auditable memory for continual interaction.
Alex: Precisely. The robot can ask itself, "Where did I last see the keys?" and trace the connections in its memory map to find the answer efficiently. It's the difference between searching a well-indexed library and rummaging through a pile of loose papers.
Sam: That makes sense for navigating a house. But what happens when the robot still gets things wrong?
Alex: This is where the system's approach to mistakes becomes interesting. The researchers call it failure-driven self-evolution. When the robot fails a task, it doesn't simply move on. It runs a diagnosis — essentially asking, "What went wrong?" Was it a bad memory lookup? A flawed plan? A tool used incorrectly?
Sam: And then what does it do with that diagnosis?
Alex: It converts the failure into what the paper calls an evo-asset. Think of it as a small, targeted patch — a new rule or a corrected procedure that gets added to the robot's internal logic. The next time it faces a similar situation, it has that patch available and is less likely to repeat the same error.
Sam: So it's building up a personal rulebook from its own mistakes, without anyone having to reprogram it from the outside?
Alex: That's a good way to put it. The robot audits its own performance, identifies what broke down, and updates its own procedures accordingly. Over time, the system becomes more reliable through experience rather than through external retraining.
Sam: How did they test whether all of this actually works?
Alex: They introduced a benchmark called EmbodiedWorldBench — a standardised set of challenges designed to push the system across sixteen different environments, ranging from indoor homes to outdoor spaces. Tasks included locating specific objects, following multi-step instructions, and navigating unfamiliar layouts.
Sam: And how did it hold up?
Alex: The paper reports that ABot-AgentOS outperformed standard baseline systems on both overall task success and on long-term information retention — meaning it was better at remembering relevant details across extended sequences of actions. The researchers are careful to frame these as promising results rather than a definitive verdict, which is appropriate given that real-world robotics testing is still limited.
Sam: It does feel like a meaningful shift in how we think about robots — less like a machine that executes a single command and more like something that actually builds up experience over time.
Alex: That's the underlying idea. And there's a practical benefit to the architecture too: by cleanly separating the thinking layer from the movement layer, you can upgrade the memory structure or the self-correction logic without having to rebuild the robot's physical control systems from scratch. Whether that holds up across a wider range of real-world conditions is still an open question, but as a direction of research, it's worth paying attention to.
Sam: A robot that learns from its own history rather than starting fresh every time — that's a genuinely different way of approaching the problem.
Alex: It is. And that's what makes this paper worth discussing. Thanks for listening to ResearchPod.