PNNL PIs: Soumya Vasisht, Wences Shaw-Cortez, Jan Drgona, Aaron Tuor, Draguna Vrabie 

University of Washington PIs: Greg Okopal, Jacob Anderson

Current state-of-the-art controllers in deployed systems lack speed, versatility, and reliable autonomy when operating autonomously in extreme dynamic environments and situations that call for adaptive responses. Modeling and control design processes must account for potential failure modes and perform fast online corrections and updates on weight-, compute-, and power-limited platforms based on noisy or partial information. Typical implementations of model predictive control (MPC) methods, while highly coveted for their favorable properties such as stability and constraint handling, can be memory and compute intensive. Novel deep-learning-based controllers mimic standard MPC behavior but suffer from inadequate constraint handling and poor performance guarantees over longer time horizons.

This project team conducts basic research to advance AI-enriched, domain-aware modeling and optimization with guaranteed outcomes for a high-consequence system use case. Our project is designed to develop and deploy a rapid, safe, robust, and adaptive deep-learning-based modeling and controller design method on the Iver2 autonomous underwater vehicle platform in collaboration with the Applied Physics Laboratory at University of Washington.

Our approach shares both the trusted autonomy of MPC and the scalability of explicit deep learning control policies. Our method, Differentiable Predictive Control (DPC), developed under PNNL’s MARS initiative, employs domain-aware deep learning to both model nonlinear dynamics and optimize explicit control policies with constraint satisfaction from small sets of experimentally recorded system measurements. The feasibility and efficiency of DPC have previously been validated both in simulations and in experiments with an embedded implementation on a laboratory device using a low-resource Raspberry-Pi platform. In this project the team will extend that outcome to a real-world autonomous vehicle platform set in a dynamic environment with safety- and mission-critical objectives and constraints. We are enhancing the DPC approach to synthesize generalizable and adaptive control policies that account for new, unseen vehicle configurations and environmental disturbances. This project is demonstrating the adaptive capabilities of DPC to model and control a generic autonomous underwater vehicle.