Reliable Deep Learning for Data-Driven Mobility Prediction under Uncertainty for Off-Road Autonomous Ground Vehicles

In the era of autonomous assets and connected vehicles, sensors mounted on off-road autonomous ground vehicles (AGV) allow us to collect large-volume vehicle mobility data (e.g. speed, fuel consumption, acceleration) from a large number of fielded vehicles under various terrain conditions. This “big mobility data” contains valuable and realistic information of vehicle mobility in various off-road environments. The huge amount of information, however, has not yet been explored in developing the Next Generation NATO Reference Mobility Model (NG-NRMM). Leveraging “big mobility data” has great potential to accelerate the development of a NG-NRMM (1) by eliminating simplifications and assumptions in simulations, and (2) by accounting for high-dimensional environmental (e.g. soil) and vehicle system parameters in a holistic manner.

The key to unlock the significant potential of “big mobility data” is big data analytics and artificial intelligence/deep learning techniques since they are flexible in “learning” a complex phenomenon using a large volume of data. Deep learning has shown substantial accuracy advantages over other predictive models. Along with the capability of connected vehicles, deep learning has opened a new venue for developing a NG-NRMM. This aligns well with the mission of both ARC and GVSC in developing cutting- edge techniques for the next-generation autonomy.

While AI has significant potential in developing such a powerful data-driven mobility predictive model, the accuracy of an AI model can be severely affected by the data quality used for training. To fully realize the potential of an AI-based mobility prediction, the following challenge must be addressed: How to guarantee the reliability of AI systems at battlefield, especially for mission-critical applications such as mobility prediction?

This research will address this challenge through uncertainty quantification (UQ) and verification and validation (V&V) of the AI models to ensure that AI-based mobility models can reliably predict vehicle mobility. Even though the proposed framework will be for mobility prediction, it is also applicable to other industries

Publications:

• Liu, Y., Jiang, C., Zhang, X., Mourelatos, Z.P., Barthlow, D., Gorsich, D., Singh, A., and Hu, Z., 2022. Reliability-Based Multivehicle Path Planning Under Uncertainty Using a Bio-Inspired Approach. ASME-Journal of Mechanical Design, 144(9), p.091701.
• Liu, Y., Barthlow, D., Mourelatos, Z.P., Zeng, J., Gorsich, D., Singh, A., and Hu, Z., 2022. Mobility Prediction of Off-Road Ground Vehicles Using a Dynamic Ensemble of NARX Models. ASME- Journal of Mechanical Design, 144(9), p.091709.
• Yin, J., Mourelatos, Z., Gorsich, D., Singh, A., Seth, T., and Hu, Z., “An Efficient Surrogate Modeling Method for Reliability-Based Global Path Planning of Off-Road Autonomous Ground Vehicles”, 2023 ASME International Design Engineering Technical Conferences & Computers and Information in Engineering Conference, Boston, MA, August 20-23, 2023.
• Liu, Y., Jiang, C., Zhang, X., Mourelatos, Z., Gorsich, D.,2022, “Simulation-Based Multi-Vehicle Path Planning Under Uncertainty Using Physarum-Based Approach”, 2022 AIAA SciTech Forum, San Diego, CA, Jan. 03-07.
• Liu, Y., Barthlow, D., Mourelatos, Z., Zeng, J., Gorsich, D., Singh, A., and Hu, Z.*, 2022, “Mobility Prediction of Off-Road Ground Vehicles Using a Dynamic Ensemble of NARX Models”, in ASME 2022 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, American Society of Mechanical Engineers, St. Louis, MO, August 14- 17.

#1.A90