The future of autonomous driving relies on AI, but it's a complex challenge. At Waymo, we've developed a unique approach to ensure safety is at the heart of our AI systems. Our mission is to create demonstrably safe AI, and we've achieved incredible results.
With over 100 million autonomous miles driven, we've made a significant impact on road safety. Our AI has reduced serious crashes by over ten times compared to human drivers. But here's where it gets controversial: we believe safety should be the foundation, not an afterthought.
Waymo's Holistic AI Approach:
Unlike other AI applications, we prioritize safety from the ground up. Our AI ecosystem consists of a Driver, Simulator, and Critic, all working together seamlessly. This holistic approach ensures that safety is proven, not just promised.
The Waymo Foundation Model is the cornerstone of our AI. It's a versatile, state-of-the-art world model that powers our entire ecosystem. By leveraging learned embeddings and supporting end-to-end signal backpropagation, we achieve powerful correctness and safety validation.
Our Think Fast and Think Slow architecture enables rapid reactions and complex semantic reasoning. The Sensor Fusion Encoder fuses camera, lidar, and radar inputs, while the Driving VLM understands rare and novel scenarios. Together, they contribute to safe and efficient driving decisions.
Distilling Knowledge:
We adapt the Waymo Foundation Model to create Teacher models for each component. These Teacher models are then distilled into smaller, efficient Student models for real-time decision-making and large-scale simulations. Distillation allows us to retain the performance of large models within more compact versions, resulting in better scaling laws.
Our Driver models generate safe and comfortable action sequences. The onboard architecture mirrors the Foundation Model, and a rigorous validation layer ensures the trajectories are safe.
The Simulator creates hyper-realistic and physically correct virtual environments for training and testing. The Critic, our evaluation system, stress-tests the Driver and identifies edge cases, enabling targeted improvements.
Continuous Improvement:
The Waymo Driver is constantly learning and evolving. Our inner learning loop utilizes Reinforcement Learning within a safe simulated environment. The outer learning loop is powered by real-world driving data, creating an even more powerful flywheel. The Critic automatically flags suboptimal behavior, and we generate improved training data from these events.
Our vast fully autonomous data is a game-changer. It enables the Waymo Driver to learn from its own experience and continuously refine its skills. There's no substitute for this real-world data, as it captures a spectrum of situations and reactions that cannot be replicated through simulation or manual driving.
Setting the Standard:
By embracing this holistic AI approach and building learning flywheels, we're not only advancing autonomous driving but also setting a new standard for safety. We're pushing the boundaries of what's possible, and there's still much exciting work ahead in the world of AI. Join us on this journey as we continue to innovate and make our roads safer.
