PhyAgentOS provides a self-evolving embodied AI operating system built on agentic workflows. It addresses the tight coupling of reasoning and execution in traditional embodied AI by introducing Cognitive-Physical Decoupling and a Session-Centered Runtime, enabling a single codebase to control diverse hardware with enhanced safety and auditability.

How It Works

The system employs Cognitive-Physical Decoupling and a Session-Centered Runtime, replacing legacy Driver-Center designs. A WatchdogSupervisor orchestrates a pipeline (SessionRunner, SkillRuntime, TargetSessionHandle). Hardware integration is simplified via Target Adapters (approx. 100 lines), decoupling core logic from specific platforms. This facilitates "One Codebase, Any Hardware" and seamless migration across simulation, game, and real-robot targets. Safety is paramount, enforced by Critic validation, Strict Preflight checks, and Target-side SafetyGuard, crucial for real-robot deployment.

Quick Start & Requirements

Clone the repo and install via pip: pip install -e . (Python ≥ 3.11). Dev dependencies: pip install -e ".[dev]". Initialize workspace with paos onboard. Run runtime with python -m PhyAgentOS.runtime.watchdog and agent with paos agent. A hardware-agnostic smoke test is available via utility scripts. Documentation links: Project Website, User Manual, Dev Guide.

Highlighted Details

• Cognitive-Physical Decoupling: Enables "One Codebase, Any Hardware" via minimal Target Adapter implementation.
• Multi-Layer Safety: Integrates Critic validation, Strict Preflight (10 checks), and Target-side SafetyGuard for secure real-robot operation.
• Fully Auditable: All states, actions, and perception results are logged to Markdown/YAML for complete traceability.
• Diverse Target Support: Compatible with game (e.g., Minecraft), debug (mock), simulation (e.g., RoboCasa), and real-robot platforms (e.g., Franka).
• Fleet Mode: Supports multi-robot coordination with shared workspaces and priority scheduling.

Maintenance & Community

Jointly developed by Sun Yat-sen University HCP Lab and Peng Cheng Laboratory. Contributions via PRs and Issues are welcomed. Development roadmap available in "Dev Plan".

Licensing & Compatibility

Released under the permissive MIT License, allowing commercial use and integration into closed-source projects.

Limitations & Caveats

Full deployment and testing on physical hardware require specific robotic platforms and configurations. The README does not explicitly detail alpha/beta status or known bugs.