Model Madness

Interactive World-Model Car Racing Duel

This project showcases an end-to-end World Model system where two AI controllers compete head-to-head in the CarRacing-v3 environment. You can watch races live, adjust high-level strategy knobs in real time, and explore how learned policies adapt.

Inspired by the World Models framework by David Ha and Jürgen Schmidhuber

---

Why This Matters

• Head-to-Head Evaluation
  Instead of evaluating agents in isolation, this duel system compares two learned controllers side by side under identical conditions.

• End-to-End World Models

  • VAE — compresses raw 96×96 RGB frames into latent vectors
  • MDN-RNN — models temporal dynamics in latent space
  • Controller — linear policy over latent + hidden states

• Interactive
  A Streamlit UI allows users and judges to adjust knobs like risk aversion or action smoothing and immediately see the difference.

• Relevant
  Competition between models is a proxy for real-world AI-vs-AI scenarios (autonomous driving, markets, security).

---

System Architecture

CarRacing-v3 (x2) --> VAE (encode frames) --> MDN-RNN (temporal model) --> Controller (policy) --> Actions

Two independent pipelines (Player A and Player B) run in parallel.
A FastAPI backend manages simulation, telemetry, and frames.
A Streamlit frontend displays the race and exposes interactive controls.

Quickstart

1. Install requirements

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Prepare weights Put trained weights in server/weights/: vae.pth
   mdnrnn.pth
   controllerA.npz
   controllerB.npz

3. Start backend uvicorn server.run_duel:app --host 0.0.0.0 --port 8000

4. Start UI BACKEND_URL=http://localhost:8000 streamlit run app.py

Controls Reset — restart duel with given/random seed Start / Pause — control simulation Per-player knobs alpha_action_smoothing risk_aversion emergency_brake_threshold lookahead_horizon gating_bias.{rain, drift, straight} Backend enforces action clamping and a minimum throttle so cars always move.

API Endpoints GET /health — check environment and weights POST /control — {action: start|pause|reset} POST /update_knobs — update controller parameters GET /telemetry — returns step count, rewards, off-track time, and frames

Repo Structure server/ run_duel.py # FastAPI backend models/ # VAE, MDN-RNN, Controller weights/ # trained weights go here configs/ bounds.yaml # knob parameter ranges app.py # Streamlit UI requirements.txt README.md