About

Motivation

This project aims to reduce design time, while enhancing the robustness, of learning enabled components embedded in autonomous systems. The challenges here are indeed DARPA-hard for the time critical and cost-effective fielding of high confidence autonomous systems for the Department of Defense and civilian sectors.

The posit of TIAMAT is that the key to the development of rapid autonomy transfer from low-fidelity simulation-only to high-fidelity simulation-to-real may not need very complex models. As statistician George E.P. Box observed, “All models are wrong, but some are useful.” It is impossible to model everything, and the fidelity of the model is driven by the complexity of the autonomy task at hand. Model parsimony is critical: Models need to be detailed enough to achieve high autonomy, but not so complex as to confuse learning. Models need to be detailed enough to achieve high performance of the autonomous system, but not be so complex as to model irrelevant statistical variations (expensive) or worse confuse (through over-fitting) learning and adaptation algorithms. Furthermore, very complex models of autonomous systems are hard to explain and collaborate with.

Technical Approach

To address this, we are developing a design methodology and tool chain suite will enable autonomy superiority for Department of Defense systems and enable more rapid certification of autonomous systems and integration in mixed initiative systems. It will impact safety critical civilian systems where autonomy solutions are fielded after “testing till the money runs out.”

Development of a tool chain suite for the design of robustly implementable autonomous systems has the following steps:

• Generating of specifications from linguistic specifications.
• Building autonomy solutions using low fidelity models, with robust adaptation and learning to account for model mismatch.
• Conducting performance assessment of the robust design methods on low-fidelity simulations.
• Determining the need for model augmentation to incorporate high-fidelity components.
• Retraining of high-fidelity models and redesign of the autonomy solution.
• Testing and evaluation of high-fidelity models in Simulation-to-Real scenarios.

Research, Development, and Experimentation

The tool chain suite will consist of Algorithms, Software, and Design Tools. The approach is broken up into two technical challenges:

Time Critical Autonomy Transfer

• Generating Specifications for Autonomy
• Robust Design Methods
• Sample Complex and Sufficiently Rich Training Data
• Design for Safe Autonomy

Model Refinement and Autonomy Redesign

• Generating New Specifications
• Model Enhancement and Training
• Efficient and Time Critical Redesign

We will also conduct Experimentation and Evaluation on a variety of platforms to assess performance and validate the algorithms, software, and design tools. These platforms include:

• Robotic Autonomous Race Cars
• Walking and Locomotion Robots
• Unmanned Aerial Vehicles and Autonomous Aircraft
• Robotic Manipulators