Gemma4Brain: core process_frame() — unified vision + nav JSON

[chatde]

Parent PRD

#1

What to build

The core of gemma4_brain.py: a process_frame(pil_image) method that makes a single Gemma 4 call and returns a navigation + scene analysis dict compatible with the existing RoverBrain/BehaviorArbiter interface. No memory injection yet — just the bare pipe working.

Acceptance criteria

• Gemma4Brain.process_frame(pil_image) calls Ollama /api/chat with image_b64 and returns a dict
• Dict has required keys: scene, nav_decision, danger_level, reaction, objects
• nav_decision is one of: FORWARD, TURN_LEFT, TURN_RIGHT, BACKUP, STOP
• danger_level is an integer 0-3
• JSON parse failures fall back to _fallback_result() (no crash)
• Gemma4Brain._check_ollama() returns True when gemma4:e4b is loaded
• BehaviorArbiter.translate_ai_decision(result) works with Gemma4Brain output (same shape as AIBrain)

Blocked by

None — can start immediately

User stories addressed

• Gemma 4 makes navigation decisions using camera vision
• Single model replaces both LLaVA and Llama 3.1