Critical driving behaviours with Driver’s Risk Field

Yurui Du1,2, Flavia Sofia Acerbo2, Jens Kober1, and Tong Duy Son2

1 Department of Cognitive Robotics, Delft University of Technology, Delft, the Netherlands
2 Siemens Digital Industries Software, Leuven, Belgium

This work uses Driver’s Risk Field to generate safety-critical simulation scenarios and augment demonstrations for data-efficient imitation learning in autonomous driving.

IFAC-PapersOnLine 2023 Imitation learning Critical scenarios Driver’s Risk Field
Paper Preprint code
Three-stage method overview: data-efficient imitation learning, critical scenario generation, and critical data augmentation

Highlights

Results

Data-efficient imitation learning improves driving behaviour.

Qualitative comparisons from the data-efficient imitation-learning results show how the spline-parameterized policy handles urban interactions compared with the Urban Driver baseline.

Ours
Baseline
Ours
Baseline
Videos from the qualitative-result examples for data-efficient imitation learning.

Generated critical scenarios expose harder failure cases.

DRF-based traffic agents generate critical interactions that reveal failures not exposed by the ordinary logged scenario. These side-by-side videos show how a critical scenario can turn a hidden weakness into an observable failure.

Critical
Log
Critical
Log
Critical-vs-log videos from the critical-scenario qualitative results.

Critical-data augmentation reduces passive driving.

Critical data augmentation targets passive-driving failure modes: before retraining, the policy drives passively; after retraining, it performs a more normal speed-up in the same type of interaction.

Before retraining
After retraining
Before/after comparison: passive driving is replaced by normal speed-up after retraining.

More qualitative examples

Videos