A 5 year journey in making autonomous systems more dependable

Over the past five years, our team at the LESS LAB has made significant strides to develop more dependable autonomous systems, particularly those relying on machine learning. Here is a summary of this effort.

  1. Domain-Specific Abstractions for Multi-Dimensional Sensor Inputs: We extract meaningful entities and relationships from raw sensor data (like images or point clouds). This allows us to: a) Generating More Realistic and Diverse Inputs, b) Specifying and Assessing Higher-Level Properties.
  1. Type Systems for Detecting World Inconsistencies: We develop type systems that catch inconsistencies between system code semantics and the physical world.
  1. Verification Frameworks for Learned Components: We create frameworks that verify DNNs against a wide range of properties, including robustness and reachability, and that focus on the input distribution to generate more useful counterexamples and be more efficient.
  1. Testing frameworks that consider the Physical State , besides the cyber state, of the autonomous system to generate more cost-effective system tests.  

This effort has helped to foster the growth of exceptional PhD students including: John Paul Ore (now at NCST), David Shriver (now at CMU-SEI), Carl Hildebrandt, Meriel Stein, Trey Woodlief, and Felipe Toledo.  I am very thankful to them and to my close collaborator, Matthew B. Dwyer, whose brilliance and kindness have served me as a reference for almost two decades.

Sebastian Elbaum
Professor of Computer Science.