Local Reasoning Session

The ATF query layer works on the local system. Operators can search the compiled wiki and the operational ledger without shipping the evidence corpus to a cloud runtime.

RobotRoss System Overview

1. Introduction

RobotRoss is an autonomous robotic artist platform designed to bridge AI creativity with physical execution. It uses a Huenit robotic arm to draw plotter-optimized SVGs or write calligraphy while providing poetic narration in the style of Bob Ross.

2. Core Architecture

The system follows a three-layer architecture:

• Commerce Layer: Commerce Layer handles order intake, queue management, and the competitive bidding system.
• Orchestration Layer: Job Orchestration (primarily bob_ross.py) manages the job lifecycle, narration, recording, and uploading.
• Hardware Layer: Hardware Interface provides low-level G-code control and calibration for the robot arm.

3. Key Subsystems

• Commerce Layer: Salesman API and order queue management.
• Job Orchestration: The brain of the artist side.
• Hardware Interface: Serial G-code control of the Huenit arm.
• Order Management: Polling and logging of drawing jobs on the Artist side.
• Narration: Local LLM-based poetic commentary.
• Video Proof: Automated OBS recording and YouTube uploading.

4. Key Integration Topics

• Bidding Rules: Competitive overwrite rules and the 8x8 Wall of Fame.
• Shopify Integration: Webhooks, metadata write-back, and human e-commerce.
• Virtuals ACP: Agentic commerce protocol for autonomous hiring.
• Calibration: Necessary startup procedures for hardware accuracy.
• Compliance: EU AI Act mapping and architectural traceability.
• Voice Control: Whisper STT, local reasoning, and Voxtral TTS.
• Architecture: ATF structure and Flotilla execution context.

5. Hardware Requirements

• Computer: Mac Mini M4 (Apple Silicon, production/Shopify pipeline) or a Windows laptop (hackathon/demo build, browser-based Control Center)
• Robot: Huenit Robotic Arm
• Cameras: Reolink 4K (Main) + macOS Screen Capture (Board) — Mac Mini build only
• Audio: BlackHole 2ch for internal routing — Mac Mini build only

6. Software Stack

• OS: macOS (darwin) for Artist (production); Windows for the Mistral-hackathon build; Ubuntu (Linux) for Salesman.
• LLM: Mistral (ministral-3:8b, local via Ollama) — default as of the 2026-07 Windows/Mistral-hackathon build. Apertus 8B remains available as an explicit fallback (--brain apertus), not auto-triggered.
• Agent Framework: OpenClaw (Mac Mini production flow); control_center.py local web UI + ATF (atf_local_server.py) for the Windows hackathon build.
• Utilities: OBS Studio, ffmpeg, Python 3.12, Node.js

7. Notes and Open Points

• Calibration Persistence: Source code indicates calibration is required after every restart (READY_FLAG in /tmp), but some docs suggest it might be semi-persistent.
• Pen Pressure: Manual leveling of the table is mentioned as a physical requirement that software cannot currently compensate for.
• Narration Latency: Narration generation now runs asynchronously in a background thread (bob_ross.py, 2026-07-09) — the arm starts drawing on generic filler commentary immediately rather than blocking on the ~60-90s local LLM call, swapping to the real narration once it lands.

---

Sources:

• AGENTS/CONTEXT/robot_ross_artist.md
• AGENTS/CONTEXT/robot_ross_salesman.md
• ~/.openclaw/workspace/skills/robot-ross/bob_ross.py