Explainable AI self-driving controller

Overview

This is the first effort to introduce explainability into a self-driving controller, by enabling it to generate textual explanations of its behavior. Specifically, we propose an introspective textual explanation model for self-driving cars to provide easy-to-interpret explanations for the behavior of a deep vehicle control network. We integrate our explanation generator with the vehicle controller by aligning their attention to ground the explanation. Ours is the first attempt to justify the decisions of a real-time deep controller through a combination of attention and natural language explanations on a stream of images.

We also provide a large-scale Berkeley DeepDrive eXplanation (BDD-X) dataset with over 6,984 video clips annotated by humans with driving descriptions, e.g., “The car slows down” and explanations, e.g., “because it is about to merge with the busy highway”. Our dataset provides a new test-bed for measuring progress towards developing explainable models for self-driving cars.

Intended Use

This software will be of interest to researchers and developers in the area of autonomous driving who are concerned with the issue of explainability. The proposed models require human ground-truth textual action description-explanation pairs, otherwise they could be applied to other scenarios and datasets.

Model/Data

An input to the model is a driving video, an output is ego-motion (acceleration, change of course) and a textual description-explanation pair.

The BDD-X dataset will be of interest to researchers and developers in the area of autonomous driving who are concerned with the issue of explainability. In particular, the collected human ground truth explanations can help inform and improve future self-driving systems and serve as a benchmark for evaluating the correctness of the generated explanations.

Limitations

The proposed models require human ground-truth textual action description-explanation pairs.

References

@inproceedings{kim2018textual,
  title={Textual explanations for self-driving vehicles},
  author={Kim, Jinkyu and Rohrbach, Anna and Darrell, Trevor and Canny, John and Akata, Zeynep},
  booktitle={Proceedings of the European conference on computer vision (ECCV)},
  pages={563--578},
  year={2018}
}