Self-supervised learning for autonomous driving

Can autonomous vehicles learn to drive safely from watching large amounts of unlabeled data? We present some of our work on self-supervised learning for scene flow, shape completion, occupancy forecasting, safety envelope estimation, and policy learning.

Learning for High Performance yet Safe Control

In-person and GatherTown

https://www.aicrowd.com/challenges/learn-to-race-autonomous-racing-virtual-challenge

Title: Reward (Mis)design for Autonomous Driving and Accumulating Safety Rules from Catastrophic Action Effects

This talk highlights two recent findings pertaining to safety of autonomous driving. The first is that most RL researchers use reward functions that are riskier than those reflected by drunk teenage drivers; the second is a method for monotonically improving safety of a fleet of vehicles over time.

Talk title: Multi-modal sensor fusion for the Amazon Scout robot

Description: I will describe the perception challenges faced by the Amazon Scout robot and our approach to doing multi-modal sensor fusion in a safe and robust way to address these challenges.

Learning How To Be Safe

In-person and GatherTown

Practical Safety Assurances for Dynamic Human-Robot Interactions

An outstanding challenge with safety methods for human-robot interaction is reducing their conservatism while maintaining robustness to variations in human behavior. In this talk, I propose that robots leverage online data of human-robot interaction to modulate the conservatism of their safety monitors, automatically shifting between robust zero-sum models to general-sum game-theoretic models of human interaction.

Title: Reinforcement Learning for Self Driving Cars

Title: Learning Plannable Representations and Planning with Learnable Skills

Abstract: Reinforcement learning provides a powerful framework for automatically learning control policies for autonomous systems. However, using RL in real-world settings, particularly complex and safety-critical domains such as autonomous driving, is often impractical due to the need for costly and dangerous exploration. In this talk, I will discuss how the framework of offline reinforcement learning can drastically expand the applicability of RL to real-world settings, discuss the fundamentals of offline RL algorithms and recent innovations, and some of our recent experience applying RL to problems in robotics and mobility.