A minimal, inspectable Physical AI Agent control plane for ListenAI ARCS Mini / LS26 — with a zero-dependency PC simulator and an evidence-backed hardware adapter boundary.

What it is. Natural-language commands become validated Action objects, pass through an allow-listed Skill Registry, and execute against a simulator or hardware adapter.
Why LS26. ARCS Mini pairs an LS2684L0U dual-core RISC-V SoC with exposed GPIO, display, camera, audio, and connectivity surfaces suitable for compact physical-agent prototypes.
Demo. The included local console shows the routed JSON, policy decision, selected Skill, simulated board state, GPIO, robot state, display, and audit log.
Truthful scope. The simulator is fully runnable on a PC. The C GPIO adapter only uses APIs confirmed in the official checkout, but has not been integrated, flashed, or tested on physical hardware.

flowchart LR
    U[Natural Language] --> R[Intent Router]
    R --> A[Structured Action]
    A --> P{Policy / Schema}
    P --> S[Skill Registry]
    S --> H[Hardware Adapter]
    H --> SIM[PC Digital Twin]
    H -. hardware validation pending .-> LS[ARCS Mini / LS26]

Python 3.8+ is the only requirement.

python -m simulator

Open http://127.0.0.1:8080 if the browser does not open automatically. Try 检查设备状态, 打开机器人任务模式, then 执行 task_001.

Run the entire host verification suite:

python scripts/check.py
python -m unittest discover -s tests -v
python -m compileall -q physical_ai_agent simulator

On systems with make, make run and make verify provide the same flows.

The five Skills are device_status, gpio_control, display_message, robot_task, and emergency_stop. See ARCHITECTURE.md for lifecycle and safety boundaries.

The upstream repository is intentionally not vendored here. Clone it beside this project or into official/arcs_mini:

git clone https://cloud.listenai.com/CSKG836746/arcs-sdk/public/arcs_mini official/arcs_mini
cd official/arcs_mini
git config core.longpaths true
.\arcs-sdk\env.ps1 check
.\arcs-sdk\env.ps1 setup
.\arcs-sdk\build.ps1 -C -S .\samples\helloworld -DBOARD=arcs_mini

The source argument above is relative to the arcs-sdk directory because the official SDK build script changes its working root. For the complete ARCS Mini voice-assistant firmware, the upstream root documents:

.\build.ps1 -S .\apps\arcs-mini -DBOARD=arcs_mini

Do not flash until you have reviewed the upstream board, flash, and recovery instructions and have the correct device attached.

Exact evidence paths are recorded in physical_ai_agent/hardware/listenai/README.md. No unconfirmed ListenAI API is called.

physical_ai_agent/
  agent/              intent_router, skill_registry, task_executor
  protocol/           action_schema, task_schema
  skills/             five uniform capability modules
  hardware/           simulator + evidence-backed ListenAI boundary
simulator/
  web/                dependency-free professional console
tests/                 router, registry, schema, executor tests
examples/              demo command catalog
docs/                  SDK evidence, demo and hardware roadmap
scripts/               dependency-free quality checks

The PC demo, protocol, agent pipeline, Skills, simulator, and host tests are complete for v0.1. The remaining hardware work is explicit: create an upstream-compatible app target, enable confirmed Kconfig symbols, integrate the adapter, compile it in that target, flash an ARCS Mini, capture logs, and validate electrical behavior. Until that happens this repository does not claim real-device execution.

EdgePilot-LS26 source is provided under the MIT License. The upstream ListenAI repository and its dependencies retain their own licenses; this repository does not redistribute them.