Adversarial Scene Generation for Virtual Validation and Testing of Off-Road Autonomous Vehicle Performance

Project started 2020.

Verifying the satisfactory operation of autonomous vehicles (AVs) is crucial to their deployment in the military context. AVs are complex machines with advanced vehicle dynamics and often consist of interconnected perception/navigation/control systems. In recent years, the algorithms that “drive” AVs are now relying more and more on deep learning; however, deep learning has been shown to be susceptible to subtle variations to its inputs further complicating the verification/testing process. Existing methods to verify, validate, and test AVs rely primarily on on-road testing typically performed in highly structured and controlled environments, but an emerging challenge is that such testing is very expensive and time-consuming. A cost-effective alternative is to use simulation, however, there only exist a few studies that attempt to identify the most difficult scene configuration using a closed-loop 3D simulation platform, and these papers do not consider the off-road setting of unstructured environments. Thus, developing methods to automatically create virtual off-road test scenarios to identify and test the limits of the capabilities of AVs in unstructured environments is an important challenge and remains an open research area.

Part of this challenge is discovering when unexpected AV behaviors arises even during normal environmental conditions, in other words how to discover the “easy” scenes that are perceived erroneously by an AV as difficult. Addressing this challenge requires not only determining what scenes are difficult for an AV to navigate, but also more importantly finding normal scenes where the AV becomes “confused,” such as in situations where straightforward paths/behaviors exist and yet the AV cannot find them and chooses a more difficult route or malfunctions (e.g., it stops).

The key idea in this project is to find the limits of deep learning algorithms on an AV by developing adversarial scenes and creating/realizing those adversarial scenes in a virtual environment where the vehicle is immersed. Such a tool to automatically construct virtual environments allows for the creation of the adversarial scenes and their application to AVs to test their behavior. This research addresses a part of a larger concerted plan to create a virtual reality lab/facility for testing autonomous vehicles.

We will focus on the following key research objectives to guide this research:

1. Create a realistic adversarial world/environment to test AV behavior and performance.
2. Construct an adversarial world that targets and challenges DNN-based perception algorithms.
3. Develop algorithms to generate perturbations (i.e., modifications to the environment) that cause large deviations in the AV decision; i.e., design adversarial environments against an AV behavior. This is especially relevant for “confusing” the AV’s autonomy stack by perturbing environments that are known to be straightforward to navigate through.

As a result, this project will enable us to automatically discover environments that are difficult for an AV to navigate through. Note that this will include the identification of environments that may be difficult for humans to navigate through and environments that are straightforward for a human to navigate through but are difficult for an AV to navigate through.

Publications:

1. T. Sender, M. Brudnak, R. Steiger, R. Vasudevan, and B. Epureanu, ”Adversarial Scene Generation for Virtual Validation of Off-Road Autonomous Vehicles,” In Proceedings of the Interservice Industry, Training, Simulation, and Education Conference (IITSEC), Orlando, Fl, Nov. 28 – Dec. 2, 2022. [PDF]
2. T. Sender, M. Brudnak, R. Steiger, R. Vasudevan, B. Epureanu, “Using Deep Reinforcement Learning to Generate Adversarial Scenarios for Off-Road Autonomous Vehicles,” In Proceedings of the Ground Vehicle Systems Engineering and Technology Symposium (GVSETS), NDIA, Novi, MI, Aug 16-18, 2022. [PDF]

References:

1. LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. “Deep learning.” Nature 521.7553 (2015): 436.
2. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., and Fergus, R. (2013). Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199.
3. Radford, Alec, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015).

#5.19