Narration Subsystem

1. Role

The Narration subsystem generates Bob Ross-style poetic commentary for each drawing job, providing a unique "personality" to the robot.

2. Key Components

• Local LLM: Mistral (ministral-3:8b) running via Ollama — the default brain as of the 2026-07 Windows/Mistral-hackathon build. Apertus 8B (Swiss open-weights, MichelRosselli/apertus:8b-instruct-2509-q4_k_m) remains available as an explicit fallback via --brain apertus, but is no longer the default and is never auto-triggered on failure.
• Voice Output: Voxtral (Mistral's TTS) by default, with Kokoro (local neural TTS) and a system-TTS fallback also selectable.
• Async generation: narration now generates in a background thread — the arm starts drawing and a generic filler intro/commentary plays immediately, then swaps to the real request-specific lines once generation finishes, instead of blocking drawing start on the ~60-90s LLM call.

3. Workflow

1. Prompt Generation: bob_ross.py constructs a prompt based on the drawing job (e.g., "a happy little lighthouse").
2. Narration JSON: the local model generates a JSON structure containing:

• intro: Spoken before drawing starts (generic filler spoken immediately; swapped for the real intro once ready, if it arrives before drawing finishes).
• commentary[]: A list of short phrases spoken during drawing at regular intervals.
• outro: Spoken after the job completes.

1. Playback: Commentary is spoken in parallel with the arm movements.

4. Notes and Open Points

• Commentary Timing: bob_ross.py uses a hard-coded COMMENTARY_INTERVAL (default 6 seconds). If the drawing is very simple, some commentary might be cut off. If it is very complex, there might be long periods of silence if the list of phrases is short.
• Concurrency: running narration generation at the same time as another local-model call (e.g. a sketch generation or an ATF chat question) can cause CPU contention on the CPU-only hackathon laptop and slow/timeout both — see the ATF fallback-chain incident note (2026-07-10 standup).
• Voice cutoff (open issue): Voxtral voice output has been observed to cut off mid-sentence during chat replies; not yet root-caused.

Sources:

• skills/robot-ross/bob_ross.py, skills/robot-ross/control_center.py, skills/voice/speak.py
• AGENTS/CONTEXT/agentegra.md
• AGENTS/CONTEXT/robot_ross_artist.md
• agentic-fleet-hub standups: 2026-07-09, 2026-07-10