Project: #405 Safety Assurance System Utilizing Visual Attention for Advanced Driver-Assistance Systems Progress Report - Reporting Period Ending: March 30, 2023 Principal Investigator: Yorie Nakahira Status: Active Start Date: July 1, 2022 End Date: June 30, 2023 Research Type: Advanced Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2022 Mobility21 UTC Progress Report (Last Updated: March 30, 2023, 11:57 a.m.) % Project Completed to Date: 75 % Grant Award Expended: 89 % Match Expended & Document: 83 USDOT Requirements Accomplishments The project's objective is to develop safe control methods for autonomous vehicles. The outcomes of the first quarter and our follow-up plans are summarized below. 1. We proposed safe control methods to ensure long-term safety probability in real-time using myopic controllers [Publication 1]. We tested the methods for driving in adversarial situations [Publication 2]. 2. We extended the methods to systems with information and communication constraints [Publication 3]. We are currently testing the methods in situations when multiple autonomous vehicles must safely drive through an intersection without traffic lights. 3. We developed tools to scale the computation of the proposed methods [Publication 4]. We plan to test this method for large-scale systems composed of many autonomous agents. 4. We proposed safe control methods for autonomous vehicles under occlusions and tested the method both in simulation and on hardware [Publication 6]. 5. We proposed a physics-informed learning framework for efficient estimation of the risk probability of stochastic systems [Publication 7]. Impacts The impacts of the project are: publications 1-7 (please see section Outputs-Publications), dissemination of results in 10 seminars or conferences; new collaborative projects: one focuses on safe autonomous driving in occluded environments and on how safe control techniques change human behaviors (please see New Partners). Other There are three major intellectual merits in the techniques we developed. First, the proposed technique can ensure long term safety using myopic controllers. There exist stringent tradeoffs between computational burden vs. long-term safety and performance, because the required computation to evaluate future trajectories often scales exponentially to the outlook horizon. Approaches based on reachability and chance-constrained optimization approaches can find safer control action in the long term, but they often come with significant computation costs. Approaches such as stochastic control barrier functions achieve a significant reduction in computational cost due to their use of myopic controllers, but can result in unsafe behaviors in a longer time horizon due to the compounding probabilities of unsafe events. The proposed technique integrates the reachability- and invariance-based (barrier-function-based) approaches to preserve the advantages of both approaches. Second, many techniques assume linear systems or systems with objectives that can be evaluated by local agents. In contrast, the proposed method allows system-wide specifications (non-decomposable barrier functions) to be ensured in nonlinear systems while admitting rigorous analysis of the cascading impact of local failures. This is enabled by our novel notion of probabilistic forward invariance/forward convergence and the resulting barrier-like function equipped with probabilistic meanings, which in turn allows the application of probability laws to convert local conditions into global guarantees. Lastly, the proposed technique avoids the common pitfalls in deterministic safe control techniques such as over-conservatism and infeasibility in large uncertainties. When deterministic methods are applied to stochastic systems, they often approximate stochastic uncertainties using bounded errors. In such settings, the bounds must be sufficiently large for uncertainties to stay within the assumed error bounds with high probability. On the other hand, some stochastic methods use probability/martingale inequalities to obtain sufficient conditions for the risk probability to stay within a tolerable level. The proposed technique is expected to avoid such pitfalls because it does not force stochastic systems into a deterministic framework, and the low computation requirements allow exact safe probabilities to be used to mediate actions within the range of affordable computation. Outcomes New Partners In our past investigation, we observed that the existing literature has the following critical drawbacks when used in autonomous driving. 1 Attempt for safer actions of the existing methods can compromise safety in the presence of strategic human behaviors. 2 The existing methods cannot deal with latent risks, such as occluded objects, which cannot be perceived and classified as threats. We are collaborating with Prof. Shirado Hirokazu (CMU) to study the solution for drawback 1. Our preliminary results have been summarized in the preprint [Publication 5]. We are collaborating with Prof. Suzuki Tatsuya (Nagoya University) to use the framework in outcome 1 to deal with the latent and invisible risks in autonomous driving. Issues None at this time.