Machine Learning Explainability and Uncertainty Quantification to Support Calibration of Trust in Automated Systems

The Explanations in Lunar Surface Exploration (ELSE) capability applies Mosaic ATMs Explainable Basis Vectors (EBV) method for explainable machine learning (xML) and likelihood scores approach to uncertainty quantification (UQ) to lunar surface exploration. In Phase I, Mosaic ATM demonstrated the ability to generalize our EBV method from discrete numerical or binary inputs (e.g., wind speed or the presence/absence of rain) to computer vision classification problems. We demonstrated the feasibility of extracting various types of information from within a deep learning model to inform qualitative and quantitative judgments of whether the machine learning (ML) model is trustworthy. Such judgments can help human and automated system users of ML model outputs decide when to trust/distrust, the systems recommendations. Such an approach to appropriately calibrate trust in automated systems is crucial to expanding their use in high risk environments like deep space exploration. In Phase II, we propose to apply the EBV method to classification of lunar terrain features to support trusted autonomy in lunar exploration, to include: Use the EBV method to produce information from within the underlying ML model to support assessment of the veracity of lunar terrain judgment model results. Incorporate EBV explanations and uncertainty quantification (UQ) into a lunar rover analog to demonstrate the ability to inform an automated system of the trustworthiness of the model. Incorporate EBV explanations into a user interface (UI) to demonstrate the ability to support appropriate calibration of human trust in an automated system. Evaluate the ELSE concept and prototype in an analog environment. We have assembled a multi-disciplinary team, partnering with the Universities Space Research Association (USRA) as a research institution and the University of Central Florida (UCF), bringing together experts in lunar exploration, ML, and human-automation interaction. Lunar surface traversability is a pressing problem in NASA space exploration that can benefit from advances in explainable machine learning (xML) and uncertainty quantification (UQ). Our Explainable Basis Vectors (EBV) method provides information from within a deep learning computer vision classification model, and our likelihood scores quantify the ability for the ML model to infer a classification from a given input image. The Explanations in Lunar Surface Exploration (ELSE) prototype will help a robot, like a lunar rover, determine whether a given region of terrain can be safely navigated. Our methodology will also allow the rover to assess the trustworthiness of the traversability judgment. Such assessments of trustworthiness will support increasingly autonomous operations, which will be required for persistent operations on the Moon and beyond. The explanations will also be interpretable by a human user, both to support remote operation of the robot and to support human assessment of the information provided by the explanation. Objectives Apply EBV Explainability and Uncertainty Quantification to Support Robotic Assessment of Lunar Terrain ML Model Judgment Integrate Explanations for Lunar Surface Explanations (ELSE) with Lunar Rover Analog Design a User Interface to Present ELSE Explanation and Uncertainty Information to a Human Evaluate ELSE Concept and Prototype in an Analog Environment Proposed Deliverables Quarterly Report #1: Including work plan Quarterly Report #2: Including ELSE ConUse and sample of labeled data Quarterly Report #3: Including information requirements for automated system determination of ML model trustworthiness and initial ML model formulation Interim Report (Quarterly Report #4): Including rover analog API requirements, UI requirements, refined model formulation, model validation results, Year 2 work plan, and Interim NTSR and NTR ELSE Software Quarterly Report #5: Including Draft Evaluation Plan Quarterly Report #6: Including Final Evaluation Plan Quarterly Report #7: Including Preliminary Evaluation Results Final Report and Briefing: Including results of operational evaluation; Phase II Project Summary, Final NTSR, NTR
Benefits: ELSE will apply our xML and UQ methods to contribute to: Successful implementation of autonomous systems to support deep space exploration, in line with efforts within Exploration Systems Development Mission Directorate (ESDMD) like Moon to Mars. Human-rover teaming in tasks involving path planning and navigation. Advances in these areas also will contribute to progress more generally in assured autonomy research, which is of interest across NASA Directorates. Non-NASA applications include robotics systems operating remotely, where increasingly autonomous operations can reduce the need for teleoperation, such as: Underground mines Radiation-contaminated sites Search and rescue in dangerous areas