Stanford Seminar - Failures & Where to Find Them: Considering Safety as a Function of Structure

Autonomous systems and robots are becoming prevalent in our everyday lives and changing the foundations of our way of life. In this talk, we’ll explore the notion of structure across high-impact application domains and consider how different contexts and tasks lend themselves to different mechanisms for safety assessment. First, we will consider reasonably structured driving environments, and explore tools for efficiently finding failures in our autonomous driving systems. Then, we’ll consider scenarios where the environment is highly unstructured and clearly defining failures becomes challenging. For such tasks, we’ll explore how anomaly detection can serve as a proxy failure identification.

Introduction.
Human-Centered Autonomy Lab.
Robust Decision-Making and Control We wish to safely and efficiently control a vehicle, despite uncertainty and disturbances.
Simulation-Based Validation.
Adaptive Stress Testing.
Multi-Agent Crosswalk Example.
Critical States . It is infeasible to examine the behavior of the system in all possible scenarios - Key Idea: Examine a smaller set of critical states that.
Classifying Critical States We seek to incorporate the notion of critical states from an onlooker's perspective into the failure search framework with human expert labels..
Data Imbalance for high-risk, low-probability events.
Multi-Task SVAE for Anomaly Detection.
Failure Identification Analysis.
Online Detection.
Anomalous Failure Detection.
Reactive Failure Detection.