PhyAgentOS is a recursive self-improvement framework for embodied agents such as robot arms and robot simulators. It connects cognitive planning, physical execution, and task verification in one feedback loop: an agent plans with a language model, acts through a physical execution layer, checks the outcome against observations and execution facts, and turns verified experience into reusable workflows and scoped Lessons for later tasks.
Physical capabilities come in Physical Execution Skills, versioned bundles that package a workflow, its required tools, and named runtime profiles for one robot or simulator. Keeping these Skills separate from planning lets the same core drive different hardware. Core bundles no model weights, simulator assets, or robot drivers. You install the Skills a deployment needs from a registry or a local bundle, start a runtime profile, and ask the agent to use it. Three control modes are supported: general-model reasoning, action models, and a hybrid of the two, depending on what the installed Skill exposes.
Verification replans within set bounds when recovery is needed, and promotion of new Lessons is guarded, with revision history. Version 1.1.0 adds centralized provider management with model discovery and per-session overrides, direct OpenAI Responses API support, and an opt-in governed interaction layer that ties tasks and sessions to multi-step decisions with independent verification.
Reported task success rates for selected configurations:
| Configuration | LIBERO-10 | RoboDojo | RoboTwin |
|---|---|---|---|
| GPT-6 + π0.5 | 100.00% | 23.30% | 76.70% |
| DeepSeek 4.1 Flash + π0.5 | 97.22% | 26.70% | 77.80% |
| GLM 5.3 Flash + π0.5 | 100.00% | 20.00% | 66.70% |
| DeepSeek 4.1 Flash | 82.00% | 8.30% | not reported |
| Kimi-K3 | 60.00% | 0.00% | not reported |
Features
- Recursive self-improvement loop: planning, physical execution, and verification feed verified experience back into Skills and Lessons
- Control modes: general-model reasoning, action models, or hybrid control
- Verified task outcomes: goals checked against observations and execution facts, with bounded replanning
- Lessons: scoped knowledge accumulated from verified episodes, with guarded promotion and revision history
- Physical Execution Skills: versioned bundles that package environment-specific tools, workflows, and runtime profiles
- Skill registry: search, install, inspect, start, and stop Skills from a registry or local bundles
- Provider management: central model configuration, connection tests, model discovery, and per-session overrides
- Governed interactions: opt-in multi-step decisions with independent root verification and deduplicated experience episodes
- Quickstarts: guides for the Piper robot arm and the LIBERO-10 simulation benchmark
