PhyAgentOS/PhyAgentOS-core

PhyAgentOS‑core – Quick explainer

What it is – An open‑source framework that lets a language‑model‑based “agent” plan tasks for a robot or simulator, run the physical actions, and then verify the outcome. The verification loop feeds back into the system so that successful experiences become reusable Skills and Lessons for future tasks. In the paper they call this Recursive Self‑Improvement (RSI).

Key pieces

Component	Role
Core Agent	Handles the chat‑style interaction with a LLM provider (OpenAI, Anthropic, etc.). Sends prompts, receives responses, and decides when to re‑plan.
Provider manager	Central config for model APIs, authentication, and per‑session overrides.
Physical Execution Skill	A packaged bundle (tar.gz) that contains the robot‑specific driver, simulation assets, and a Skill Runtime that actually moves the robot.
Forge / Tool API	Standardised calls that the Skill runtime must implement (query, action, verification).
Governed interaction (optional)	Adds a “Supervisor” that records episodes, deduplicates experience, and promotes verified lessons automatically.
Dora CLI	Runtime manager that starts/stops Skill processes and routes messages between the Core and the Skill.

How you use it

1. Install the Python package (pip install -e .).
2. Configure a model provider (e.g., paos provider configure openrouter).
3. Install a Skill for your robot or simulator (paos skill install <name>). Skills are versioned bundles that declare required tools and runtime profiles.
4. Start the Skill runtime (paos skill start …).
5. Talk to the agent (paos agent -m "Pick up the red block"). The agent plans, sends the plan to the Skill via the Forge API, the robot executes, and the system checks sensor data to verify success. Successful runs are stored as reusable lessons.

Why it matters

• Unified interface – One CLI works for any robot or simulator as long as a Skill exists.
• Verified outcomes – The system can automatically re‑plan if a task fails, rather than blindly trusting the model.
• Experience reuse – Successful episodes are turned into reusable workflows, enabling the system to improve over time (the “self‑improving” claim).
• Modular – Cognitive planning stays separate from low‑level robot drivers, making it easier to swap hardware or simulation environments.

Benchmarks – The README shows results on three standard robot‑learning suites (LIBERO‑Long, RoboDojo, RoboTwin). With a GPT‑6 model plus a “π0.5” verification layer they report up to 100 % success on LIBERO‑10 and 80 % on RoboTwin, outperforming public leaderboard baselines.

Who it’s for – Researchers building embodied AI agents, robotics labs that want a plug‑and‑play LLM‑driven control stack, and developers looking to experiment with self‑improving robot software without writing low‑level integration code from scratch.

Status – Latest stable release v1.1.0 (Oct 2026). MIT‑licensed, with docs, quick‑start guides, and a Discord community.