Sampling-Based and Gradient-Based Efficient Scenario Generation

Vidisha Kudalkar, Navid Hashemi, Shilpa Mukhopadhyay, Swapnil Mallick, Christof J. Budnik, Parinitha Nagaraja, Jyotirmoy V. Deshmukh

Abstract: Safety-critical autonomous systems often operate in highly uncertain environments. These environments consist of many agents, some of which are being designed, and some represent the uncertain aspects of the environment. Testing autonomous systems requires generating diverse scenarios. However, the space of scenarios is very large, and many scenarios do not represent edge cases of the system. We want to develop a framework for automatically generating interesting scenarios. We propose to describe scenarios using a formal language. We show how we can extract interesting scenarios using scenario specifications from sampling-based approaches for scenario generation. We also introduce another technique for edge-case scenario generation using the gradient computation over STL. We demonstrate the capability of our framework in scenario generation in two case studies of autonomous systems involving the autonomous driving domain and the safety of human-robot systems in an industrial manufacturing context.

External IDs:dblp:conf/rv/KudalkarHMMBND24