Machine Learning Models for Autonomous Vehicles


(T) Waymo’s principal scientist Drago Anguelov recently highlighted the latest innovations of Waymo’s machine learning models for its autonomous vehicles at CVPR 2020. In a previous blog article, Waymo (ChauffeurNet) versus Telsa (HydraNet)), I described Waymo’s data pipelines that are architected with four building blocks:. 

  • Perception: Find road paths, traffic lights, obstacles…Leveraged network architecture search (NAS) to find quickly best architectures for model
  • Behavior Prediction: Leverage Google maps; train agents through ChauffeurNet to estimate trajectories in a simulated environment\
  • Planning: Generate trajectories through ChauffeurNet based on feasibility, staying on the road, and avoiding collision
  • Controls optimizer: Throttle and steering

Models for perception and behavior predictions are built with supervised or self-supervised learning:

Turing award’s Professor Yann LeCunn is a strong advocate of self-supervised learning, see for more about the case that he made about self-supervised learning for autonomous vehicles in my previous blog article “the power and limits of deep learning“. Simply said, a self-supervised learning system learns by itself the labels from the data itself or more precisely, it is able to make predictions in a supervised way from acknowledging that it does not know part of the input data.

Models for planning are built with reinforcement and imitation learning:

Waymo bought end of last year, a spin-off from Oxford University Latent Logic that specialized in imitation learning for autonomous vehicles. Imitation learning enables expert demonstrations, which are often supplied from human experts themselves, to generate the training data for the system. There are two main categories of imitation learning algorithms: behavior learning that learns in a supervised way the policy from the expert demonstration, and inverse reinforcement learning that learns from the environment of the expert demonstration the reward function, and then find the optimal policy that maximizes that policy. Two good tutorials on imitation learning: an ICML 2018 Tutorial, and a lecture from UC Berkeley’s CS Professor Sergey Levine.



To learn more about the machine learning systems for autonomous vehicles, see another excellent presentation of Drago in the class of MIT Professor Lex Fridman for self driving cars:



Note: The picture above is Mori Point in Pacifica.

Copyright © 2005-2020 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com.