What it is
A robotic agent OS layer for long-horizon embodied tasks, with verification-aware execution and persistent multimodal graph memory plus a new executable benchmark. Instead of a monolithic VLM/VLA controller, it adds an Agent OS above robot skills: scene-conditioned planning, isolated skill execution, multi-stage verification, and typed multimodal graph memory with failure-driven self-evolution.
Key results
- LoCoMo: 87.5 static, improved to 88.7 with self-evolution
- OpenEQA EM-EQA: 59.9 static, improved to 60.4 with self-evolution
- Mem-Gallery: 88.6 static, improved to 89.0 with self-evolution
Shows that a separate agent OS and persistent source-grounded memory can improve long-horizon embodied behavior and continual memory quality.