metadata
title: Robust World Model Planner
emoji: ⚡
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
license: mit
short_description: Interactive world model planning simulation
⚡ Robust World Model Planner
An interactive visualization comparing different world model approaches for robot navigation and planning.
What Is This?
This demo illustrates concepts from the paper "Closing the Train-Test Gap for Gradient-Based World Model Planning". It shows how different world model training strategies affect an agent's ability to navigate obstacle-rich environments.
The Three Models
| Model | Description | Color |
|---|---|---|
| STANDARD_DINO | Baseline frozen encoder - noisy gradients | 🟣 Fuchsia |
| ONLINE_WM | Online fine-tuning - moderate improvement | 🔵 Blue |
| ADVERSARIAL_WM | Adversarial training - smooth, robust | 🔵 Cyan |
Key Concepts
- Steering Gain: How strongly the agent avoids obstacles
- Scan Horizon: How far ahead the agent "sees" (lookahead radius)
- Loss Surface: Adversarial training smooths the optimization landscape
- Gradient Field: Visualize the planning gradients across the map
How To Use
- Select a Model from the sidebar (try Adversarial for best results)
- Choose an Environment (Simple → Wall → Complex)
- Click "Execute" to start navigation
- Click anywhere on the map to set a new target
- Toggle "Matrix_View" to see the gradient field
- Explore tabs: Live Feed, Data Surface (loss landscape), Analytics (benchmarks)
Why Adversarial?
Standard world models have noisy, non-convex loss surfaces that make gradient-based planning unreliable. Adversarial training:
- ✅ Smooths the loss landscape
- ✅ Provides consistent gradients
- ✅ Enables reliable obstacle avoidance
- ✅ Closes the train-test gap
Based on research: Closing the Train-Test Gap | Built with React + Recharts