SBIR-STTR Award

Verification for deep neural networks and learning-enabled autonomous systems is critical for the safe deployment of artificial intelligence in the wild. Although significant efforts have been made in academia and industry with rigorous techniques and tools developed in this area, several grand challenges still need to be addressed for the applicability and usage of state-of-the-art techniques in real-world applications. optoXense is teaming with University of Nebraska - Lincoln to propose solutions for five key challenges in both open-loop and closed-loop verification, including 1) scalability, conservativeness, memory, and timing performance, 2) quantitative real-time verification, 3) intuitive user-friendly interface, and ROS-integrability, 4) handling temporal time-dependent properties, and 5) all-in-one demand. In Phase I, we aim to develop basic front-end and selected back-end features of the proposed tool to demonstrate feasibility. Dr. Tran of UNL has been developing the NNV (Neural Network Verification) framework for Deep Neural Networks and Learning-enabled CPS (Cited in N231-061 topic reference no. 9). The NNV tool is gaining great attention from industry and academia. It has been used for the DARPA Assured Autonomy project and is currently partially supported by Toyota Research and NSF. This proposal aims to improve on NNV by eliminating dependency on proprietary software. We aspire to deliver a tool that could become the dominant platform for neural network verification. This aspiration could be realized by a flexible architecture that could be extended by a community of fellow researchers and developers.

Benefit:
Many existing AI verification tools only work with one specific verification approach, to complement the authors papers. It is not realistic for system engineers to utilize such tools, each requiring its own interface. Our proposed tool will help consolidate interfaces, whether to verify at the individual component property level or at the system level for closed-loop intelligent autonomous systems. We intend to release the work product as open-source tool in conjunction with journal papers and conference presentations. The low barrier of adoption would accelerate adoption, providing feedback from early adopters to drive rapid feature enhancement. Since such a universal tool does not yet exist, there is opportunity for our proposed tool to become the dominant AI verification platform. An open-source release could be a powerful means to extract contribution from broader community. Many interfaces will be developed by many third-party contributors to support integration with many more tools and information systems, greatly increasing its utility. We expect developers of future AI verification algorithms will want to integrate into our platform, furthering its dominance. While UNL focuses on fundamental research, optoXense will investigate use cases for government, defense and aerospace corporations through additional contracts. The attached letters of support show such prospect in Phase II and beyond. We are already developing an automation tool to enable model-based system engineering (MBSE) for safety critical autonomous system in an ongoing Phase II SBIR. Verification for AI based perception would be a feature extension appreciated by many target users. Grandview Research estimated the global autonomous vehicles market demand at 51.6 thousand units in 2021 and is expected to expand at a compound annual growth rate (CAGR) of 53.6% from 2022 to 2030. Insider Intelligence predicts total global shipments of enterprise drone or unmanned aerial vehicles to reach 2.4 million in 2023 increasing at a 66.8% compound annual growth rate (CAGR). The operators of these enterprise applications are motivated to adopt AI verification tool to reduce its risk and liability. We expect to develop a service business for commercial adopters that require training, support, customizations, and extensions.

Keywords:
verification, verification, Computer Vision, Artificial Intelligence, autonomous control, Neural network, cyber-physical system, Digital Engineering, verification and validation