Autonomy and machine learning pose challenges for traditional, widely accepted techniques for engineering safety and reliability. However, industry and NASA depend on autonomy for mission-critical functions with a high degree of assurance. While R&D has demonstrated that autonomous robotics can accomplish amazing feats in practice, high levels of autonomous capability cannot be fully utilized in mission-critical situations due to a lack of assurance that these capabilities will be safe, reliable, and trustworthy when called upon. Fortunately, the significant economic opportunity presented by autonomous mobility, along with a promise of widespread potential safety benefits, are driving the automotive industry to address assurance challenges of autonomy. Edge Case Research is at the forefront of this revolution in safer autonomy with the release of the UL 4600 standard, the world’s first standard for evaluating the safety of autonomous products. UL 4600 addresses the need for novel technical and safety standard approaches to accommodate autonomy. This includes dealing with AI and machine learning, as well as helping to ensure the safety of vehicles that do not have a driver to handle unusual situations and equipment failures. We are pleased to propose this NASA STTR project in partnership with Carnegie Mellon University. On this project we will: (1) Identify visual perception functions that are relevant to NASA mission concepts and then draft assurance cases for them, (2) Select a visual perception function that relies on deep learning and develop a detailed validation plan for it, and (3) Demonstrate the technical feasibility this validation plan in order to: (a) demonstrate the technical viability of COTS products to assist NASA in conducting similar validation plans in the future, (b) characterize the types and volumes of test data that must be collected, and (c) explore the suitability of the validation plan results in the context of the assurance case.
Perception algorithms must be reliable, trustworthy, and robust to support a variety of NASA mission concepts. The failure of obstacle detection, localization, and mapping algorithms could put future missions and, someday, the lives of astronauts at risk. Technologies and standards for validating that these risks have been mitigated is therefore paramount for numerous future NASA applications. This project will adapt promising techniques from the automotive industry, such as the new UL 4600 standard, to meet this challenge.
The value of feasible autonomy validation to industry is hard to overstate. A lack of viable V&V strategies is among the biggest roadblocks to growth for burgeoning industries involving unmanned aerial vehicles and self-driving cars. Despite overwhelming commercial interest for autonomous vehicles, legitimate concerns about how to regulate their testing limits how these vehicles can be deployed.