The System Behind Self-Driving: Waymo’s Dmitri Dolgov

Waymo has moved from a research lab to a global service, delivering roughly half a million fully autonomous rides each week in several U.S. cities. This scale‑up reflects a shift from pure scientific exploration to real‑world deployment, supported by a robust sensor stack that fuses LiDAR, radar and high‑resolution cameras to achieve 360‑degree awareness. The hardware‑intensive perception system feeds an AI pipeline that processes data locally, while non‑critical tasks such as post‑ride cleaning alerts run in the cloud, keeping latency low for safety‑critical decisions.

At the heart of Waymo’s technology lies a large foundation model trained off‑board, which is then specialized into three teacher models: the driver, the simulator, and the critic. These teachers are distilled into smaller, real‑time inference models that run on the vehicle. This architecture balances end‑to‑end learning with structured intermediate representations, allowing the system to predict object behavior, generate plausible future scenarios, and evaluate actions against safety criteria. The simulator creates synthetic worlds for massive data generation, while the critic flags risky events, providing a feedback loop that refines the driver model without relying solely on raw pixel‑to‑trajectory pipelines.

Training combines imitation learning, reinforcement‑learning‑based fine‑tuning (RLFT), and extensive simulation to cover long‑tail edge cases. Waymo optimizes reward functions that emphasize superhuman safety, smooth acceleration, and predictable social driving, ensuring the autonomous vehicle behaves like a courteous human driver. These layered approaches explain why driver‑assist technologies cannot simply evolve into full autonomy; they lack the comprehensive simulation, critic feedback, and safety‑first reward structures that Waymo has built. As compute power and AI methods continue to advance, Waymo’s model‑centric, simulation‑driven strategy positions it to expand globally while maintaining rigorous safety standards.