Explore an invited talk from the GPU Technology Conference (GTC) 2018 that delves into learning steering bounds for parallel autonomy and handling ambiguity in end-to-end driving. Discover the latest methodologies for training end-to-end systems in autonomous vehicles, focusing on the challenges of integrating decision-making capabilities beyond reactionary control. Examine the concept of parallel autonomy, its architecture, and hardware implementation. Investigate approaches to learning steering distributions, including discrete action learning and multimodal distributions. Gain insights into dataset collection, Bayesian deep learning, and uncertainty estimation in end-to-end steering control. Analyze the advantages and limitations of these approaches in the context of autonomous driving pipelines and higher-level decision making.
Learning Steering for Parallel Autonomy
Alexander Amini and Massachusetts Institute of Technology via YouTube
Overview
Syllabus
Intro
Motivation
Autonomous Driving Pipeline
End-to-End Learning
Challenges
Talk Outline
Guardian Angel
Parallel Autonomy:Architecture
Parallel Autonomy: Hardware
Shared # Binary Control
Possible Approaches
Autonomous Modes
Related Work End-co-End Learning
Learning a Steering Distribution
Discrete Action Learning
Multimodal Distributions
Advantages of this approach
Dataset Collection
Discrete to Continuous
Variational Bayes Mixture Models
Bounds for Parallel Autonomy
Why Care About Uncertainty?
Bayesian Deep Learning
End to End Steering Control
Integrating Uncertainty Estimation
A Bayesian Outlook on End to End Control
Elementwise Dropout for Uncertainty
Spatial Dropout for Uncertainty
Training Results
Summary
Taught by
https://www.youtube.com/@AAmini/videos
