Learning Resource for AV Simulation

Master Waymax & JAX

A hands-on learning project to understand Waymo's high-performance autonomous driving simulator. Explore data-driven simulation, JAX optimization, and sim agent research.

1000+
Hz Simulation Speed
100K+
Real Scenarios
8
Deep Dive Papers
12K+
Lines of Docs

Why Learn Waymax?

Waymax achieves a breakthrough by combining real-world data, GPU acceleration, and multi-agent simulation—100x faster than traditional simulators while maintaining realism.

Data-Driven Simulation

Learn how Waymax uses real Waymo Open Motion Dataset (100K+ scenarios) instead of synthetic environments for realistic simulation.

JAX-Powered Performance

Achieve 1000+ Hz simulation speeds using JAX's JIT compilation, vmap vectorization, and GPU/TPU acceleration.

Multi-Agent Simulation

Simulate all agents in a scene, not just the ego vehicle. Train realistic sim agents that match human driving behavior.

Closed-Loop Evaluation

Understand why closed-loop testing is critical for AV validation and how to properly evaluate planning systems.

Structured Learning Path

Progress from JAX fundamentals to advanced simulation techniques with our structured curriculum.

Weeks 1-2

JAX Fundamentals

  • Arrays & Immutability
  • Automatic Differentiation
  • JIT Compilation
  • PyTrees
Weeks 3-4

Waymax Basics

  • Data Loading
  • Metrics System
  • Log Playback
  • Agent Interfaces
Weeks 5-6

Multi-Agent Simulation

  • IDM Agents
  • Vectorization
  • Batch Simulation
  • Performance Tuning
Weeks 7-12

Advanced Topics

  • RL Training
  • Behavior Cloning
  • Scenario Generation
  • WOSAC Evaluation

Key Insights

Critical lessons from studying Waymo's simulation infrastructure

Speed vs. Realism Solved

Waymax achieves 1000+ Hz on GPU with real-world data—100x faster than CARLA while maintaining realism through data-driven approach.

JAX is Ideal for Simulation

Functional paradigm, XLA compilation, and vmap enable massive parallelization that traditional OOP simulators cannot match.

Closed-Loop is Critical

Open-loop metrics don't correlate with real-world performance. Error compounding in closed-loop evaluation reveals true system capabilities.

Realism is Distribution Matching

Success means matching the distribution of human behavior, not just avoiding collisions. WOSAC metrics capture this nuance.

Ready to Start Learning?

Dive into the deep dive papers to understand the theory, explore the hands-on code examples, and prepare for ML infrastructure engineering roles.