Trajectory Planning with Omni-Experiential Learning for Robust Fleet Mobility on Extreme Off-Road Terrain

Project #5.26 begins 2026.

This project aims for robust and adaptive offroad mobility of autonomous vehicle fleets on challenging off-road terrain through a distributed, peer-based learning framework. The planned learning framework leverages experience, either of the ego vehicle or of peers, to improve mobility. It acts through a channel separate from existing learning methods and has a distinctive capability to compensate for a multitude of error-inducing physical phenomena (i.e., adapting to “unknown unknowns”). Uniquely, it is also intended to consume experience from a variety of sources, with automatic adaptation to the information quality of the source. Sources may include both physically similar and dissimilar vehicles in a heterogeneous fleet, and both high and low fidelity digital twins.

This project targets applications where autonomous fleet mobility is required in areas with extreme off-road terrain. The context of interest includes hazardous terrain topology, such as rough ground and steep slopes; uncertain and unmapped areas, such as heavily vegetated areas that restrict sensor visibility; and locomotion involving complex physical phenomena, such as tire-terrain interaction in non-homogeneous mediums or soft soils.

The state-of-the-art lacks a learning framework capable of adapting in real-time to a multitude of error-inducing physical phenomena, learning from a variety of sources, including dissimilar ones, and meeting the requirements of extreme mobility applications. Filling this gap will enable robust and adaptive off-road fleet mobility in extreme environments.

This research seeks a new paradigm for peer-based learning for local trajectory planning and control of a heterogeneous fleet of vehicles. The focus is on two fundamental research questions:

(RQ1) How can peer data strengthen the robustness of an off-road trajectory planner in real time? This research question is novel because existing approaches, as reviewed prior, either do not provide robustness to unanticipated phenomena (i.e., they only adapt to “known unknowns”), are computationally inappropriate for extreme mobility applications, or are adhoc and lack theoretical backing. Answering this question will create a fundamentally new paradigm for learning-based trajectory planning and control, and it will be tested in one of the most extreme applications available.

(RQ2) How can data-driven functional mappings increase the utility of data from dissimilar peers? This research question is novel because while map creation and (terrain) parameter identification approaches naturally extend to multiple vehicles due to the separation of the learned data from the vehicle, the same is not true of policy learning methods. Prior work has been restricted to single-vehicle applications, and extensions to multi-vehicle applications are currently only possible if the vehicles are identical. Answering this question is required to bring the benefits of the approach to heterogeneous fleet applications.

5.26