Hybrid Learning-based Policy for High-Speed Autonomous Off-Road Navigation

Project begin 2024.

The overarching goal is to develop a hybrid policy that will enable unmanned autonomous vehicles to adapt to dynamic off-road environments and extend the operational design domains (ODD). The envisioned autonomy stack should be able to use data collected online to compute trajectories that safely navigate over an unknown terrain without operator intervention. Overall, this project aims to address the following question: How can a hybrid policy be updated to account for new data collected from the environment while safely adhering to probabilistic constraints on the system dynamics?

To construct the hybrid policy, we leverage an off-policy reinforcement learning (RL) algorithm (Task 1) that optimizes the policy using an ensemble of experts (Task 2) composed of both physics-based and data-driven models and explores additional states using a risk-aware safety layer that incorporates reachability analysis (Task 3). The key objectives are:

1. Set up an off-policy RL framework capable of using data collected both online and offline.
2. Develop a hierarchical architecture that learns a modular set of behavior sub-policies using an ensemble of physics-based and data-driven models.
3. Develop a reachability-based planning layer that can adjust high-risk trajectories and enable safe exploration of new policies.

The proposed framework facilitates training with data collected online from simulations and with expert demonstration data collected offline from field tests of the ARC MRZR Vehicle Platforms.

#1.42