Learning Robust Policies for Self-Driving

Explore a comprehensive talk from the ECCV 2022 workshop on Autonomous Vehicle Vision, focusing on robust policies for self-driving vehicles. Delve into three key projects: TransFuser, PlanT, and KING. Examine the challenges of geometric fusion, the benefits of global context in perception, and innovative approaches to imitation learning. Investigate the architecture, loss functions, and experimental results of TransFuser, including its performance on benchmarks and the CARLA Leaderboard. Learn about end-to-end driving policies, intermediate representations in learned planners, and the concept of critical scenarios as attacks in autonomous driving. Gain insights into gradient paths, cost functions, and collision types in adversarial scenarios, concluding with a summary of cutting-edge advancements in self-driving technology.

Intro
Self-Driving - A Human Dream
Imitation Learning
Sensors
Geometric Fusion Lacks Global Context
TransFuser: Key Idea
TransFuser. Full Architecture
TransFuser: Loss Functions
TransFuser. Experimental Evaluation
TransFuser: Results on Longest6 Benchmark
TransFuser. Results on CARLA Leaderboard
TransFuser. Attention Maps
TransFuser. Failures
Rule-based Planner
End-to-End Driving Policy
Learned Planner
PlanT: Intermediate Representation
PlanT: Model
PlanT: Input Representation
KING: Critical Scenarios as Attacks
KING: Overview
KING: Gradient Paths
KING: Cost Functions
KING: Collision Types
KING: Adversarial Cut-in Maneuver
Summary