New frontiers in data-driven autonomous driving
CVPR tutorial

Welcome to the self-driving tutorial!

This hands-on tutorial is prepared for ML community and focused on the latest, data-driven approaches to prediction, planning, and simulation in self-driving, as well as their interplay with computer vision. You will learn basic concepts, state-of-the-art solutions and also build your own solutions.

This tutorial is prepared by Toyota Woven Planet team with the purpose of advancing state-of-the-art self-driving research. You can watch the videos here or watch the YouTube stream link: https://youtu.be/5hukp-xMc0g

Also check our published papers, datasets and competitions on this topic.

Autonomy 2.0

Modern production self-driving systems used in the industry still rely excessively on hand-engineering, especially when it comes to planning and simulation. This is becoming a limiting factor in self-driving development. Autonomy 2.0 we cover in this tutorial is a paradigm of using ML-first approaches for these components offering greater scalability, safety and comfort of self-driving cars.

Modern self-driving pipeline and the amount of ML employed in each component. The goal of Autonomy 2.0 is to make the entire stack ML-first and data-driven.

Introduction

This hands-on tutorial focused on the latest, data-driven approaches to prediction, planning, and simulation in self-driving, as well as their interplay with computer vision. You will learn basic concepts, state-of-the-art solutions and also build your own solutions. This tutorial is prepared by Lyft Level 5 team with the purpose of advancing state-of-the-art self-driving research.

Autonomy 2.0

Autonomy 2.0 is a paradigm of using ML-first approaches for building self-driving stack offering greater scalability, safety and comfort of self-driving cars.

Datasets for training

To train and test the models, you will need large-scale datasets. These are of two types: perception and prediction datasets.

L5Kit: Software development kit

Part of this tutorial is the ability to try the discussed topics and build their own ML prediction, planning and simulation modules for self-driving cars. This training is done using L5Kit - a software development platform you can find at l5kit.org

ML Perception

Perception is solving the problem of what is around the self-driving vehicle. In this part, you will learn about the state of the approaches to perception and how to train your own ML model.

Large-scale mapping

Maps allow the self-driving vehicle to pre-compute information about the environment it is operating and this way help improve perception and other tasks. In this part, you will learn how semantic mapping is done autonomously at scale.

ML prediction

Prediction is solving the problem of what are other traffic participants (vehicles, cyclists, pedestrians) likely to do next. In this part, you will learn about the state of the approaches to ML prediction and how to train your own ML model using L5Kit.

ML planning

Planning is solving the problem of what the self-driving car should do. In this part, you will learn about the state of the approaches to ML planning and how to train your own ML model using L5Kit.

Deployment & road testing

Deployment of self-driving vehicles and road testing is the final part of the development process. In this section you will learn how this part works.

ML simulation

Simulation is solving the problem of evaluating system performance offline without any road testing. In this part, you will learn about the state of the approaches to ML simulation and how to train your own ML model using L5Kit.

Scaling data

To train ML models for Autonomy 2.0 we need large-scale datasets. In this part, you learn what data do we need and where they might be coming from.

Benchmarks & competitions

As you developed your own models, you can submit them to ML competitions and compare them with others. In this section, you learn how.


Organising team