Unsupervised Extraction of Test Scenarios from Time-Series Sensor Data using Trace Graphs

Abstract

The dependability of autonomous agents, such as self-driving cars, in complex and unpredictable real-world environments is a critical challenge. To address this, scenario-based testing attempts to assess agents across a range of diverse scenarios. However, the manual definition of these test scenarios is often labor-intensive and requires considerable domain expertise, while existing methods to extract scenarios from sensor data also depend on intricate assumptions about the data. To overcome these limitations, we introduce an unsupervised approach that uses trace graphs to determine meaningful scenario boundaries in time-series sensor data, without the need for additional domain knowledge or manual input. Our experimental results demonstrate that our method can extract a small but comprehensive set of test scenarios that captures the full spectrum of the agent's experiences as observed in the sensor data.