GeminiFleet

## What it does GeminiFleet runs a physics-based warehouse simulation where autonomous robots pick up and deliver items. A fleet manager controls robot behavior through natural language — no code, no config files. **Example commands:** - "Make robots more cautious" → speed drops, safety margins increase - "Speed things up, we're behind schedule" → max speed, tighter margins - "Focus on the north side" → robots prioritize north-zone tasks Google Gemini interprets each command with full context (fleet status, delivery counts, collision stats) and generates precise parameter updates that modify robot behavior in real-time. ## How it works **PyBullet Physics Engine** — Real rigid-body simulation with collision detection. Warehouse environment with walls, shelves, pickup/dropoff zones, and 4-6 autonomous robots navigating with priority-based collision avoidance. **Gemini 2.0 Flash Policy Engine** — Translates natural language into 7 tunable parameters: speed, safety margin, congestion response, task selection strategy, cooperation mode, zone preference, and concurrency. Values are clamped to safe ranges. **Live Web Dashboard** — Real-time 2D visualization via WebSocket at 10Hz. Tracks robot positions, planned paths, carrying status, and delivery statistics. Collapsible panels for robot status and active policy display. ## Key Innovation Robot fleet behavior is parameterized into meaningful dimensions that an LLM can reliably map from ambiguous human instructions. Operational expertise — not programming skill — drives fleet optimization. ## Deployment Runs entirely on **Vultr non-GPU VMs** via Docker. PyBullet operates in CPU-only mode. Single `docker compose up` deploys the full simulation + dashboard + Gemini chat. ## Built with - **PyBullet** — Bullet Physics simulation - **Google Gemini 2.0 Flash** — NL→policy translation - **FastAPI + WebSocket** — Real-time state streaming - **Docker** — Vultr deployment