Long-horizon embodied agents often have capable perception and policy modules but lack a general runtime layer that coordinates reasoning, memory, verification, and cross-embodiment execution. ABot-AgentOS proposes treating that runtime as an explicit Agent OS: a deliberative layer above low-level controllers that unifies planning, auditable persistent memory, and gated continuous improvement so robots can execute complex, multi-stage tasks without leaking future-ground-truth during training.
Key Findings
- EmbodiedWorldBench: a new executable benchmark with 16 indoor/outdoor/hybrid scenes, four difficulty levels and 200+ tasks (navigation, object search, NPC dialogue, dynamic events) to evaluate long-horizon embodied execution and trace-grounded scoring.
- Universal Multi-modal Graph Memory: converts dialogue, visual observations, spatial context, temporal relations and task traces into typed nodes/edges to provide a persistent, source-grounded substrate for planning and verification.
- Failure-driven self-evolution loop: diagnosed memory failures are converted into gated runtime "evo-assets" that are only promoted to later evaluation splits, enabling continual improvement while preventing current-split ground-truth leakage.
- Empirical gains: ABot-AgentOS Static posts high memory scores (LoCoMo 87.5, OpenEQA EM-EQA 59.9, Mem-Gallery 88.6, NExT-QA Acc@All 76.5); self-evolution further improves LoCoMo to 88.7, OpenEQA to 60.4, and Mem-Gallery to 89.0, while improving task success and goal completion over a single-controller baseline.
Who it's for and trade-offs
Great fit if you build or evaluate long-horizon embodied agents that need persistent, auditable memory and staged verification (research labs, robotics teams, benchmarkers). It is practical for setups that can integrate a deliberative layer above existing controllers and accept extra runtime complexity for memory management and edge–cloud collaboration.
Look elsewhere if your focus is purely on low-level control optimization, millisecond-scale reactive policies, or very lightweight embedded platforms where the overhead of a multi-modal persistent memory and verification pipeline is infeasible.
Where it fits
ABot-AgentOS targets the gap between high-quality perception/policy models (VLMs/VLAs) and deployed robots: instead of collapsing decision logic into a single controller, it provides a modular OS-like runtime for planning, tool use, verification, and continuous, auditable memory—making it suitable as the middleware for embodied AI stacks and multi-controller systems.
How it works (brief)
The system exposes scene-conditioned planners, context-isolated skill executors, multi-stage verification routines, and a Universal Multi-modal Graph Memory that records observations and traces with provenance. The failure-driven self-evolution process detects memory-related failure modes, wraps fixes as gated artefacts, and only promotes them to future evaluation splits to avoid leaking current-split ground truth while enabling continual improvements.