Abstract:Specialized machine learning models, regardless of architecture and training, are susceptible to failures in deployment. With their increasing use in high risk situations, the ability to audit these models by determining their operational design domain (ODD) is crucial in ensuring safety and compliance. However, given the high-dimensional input spaces, this process often requires significant human resources and domain expertise. To alleviate this, we introduce \coolname, an LLM-Agent centric framework for automated generation of semantically relevant test cases to search for failure modes in specialized black-box models. By leveraging LLM-Agents as tool orchestrators, we aim to fit a uncertainty-aware failure distribution model on a learned text-embedding manifold by projecting the high-dimension input space to low-dimension text-embedding latent space. The LLM-Agent is tasked with iteratively building the failure landscape by leveraging tools for generating test-cases to probe the model-under-test (MUT) and recording the response. The agent also guides the search using tools to probe uncertainty estimate on the low dimensional manifold. We demonstrate this process in a simple case using models trained with missing digits on the MNIST dataset and in the real world setting of vision-based intruder detection for aerial vehicles.
Abstract:Observing and filming a group of moving actors with a team of aerial robots is a challenging problem that combines elements of multi-robot coordination, coverage, and view planning. A single camera may observe multiple actors at once, and the robot team may observe individual actors from multiple views. As actors move about, groups may split, merge, and reform, and robots filming these actors should be able to adapt smoothly to such changes in actor formations. Rather than adopt an approach based on explicit formations or assignments, we propose an approach based on optimizing views directly. We model actors as moving polyhedra and compute approximate pixel densities for each face and camera view. Then, we propose an objective that exhibits diminishing returns as pixel densities increase from repeated observation. This gives rise to a multi-robot perception planning problem which we solve via a combination of value iteration and greedy submodular maximization. %using a combination of value iteration to optimize views for individual robots and sequential submodular maximization methods to coordinate the team. We evaluate our approach on challenging scenarios modeled after various kinds of social behaviors and featuring different numbers of robots and actors and observe that robot assignments and formations arise implicitly based on the movements of groups of actors. Simulation results demonstrate that our approach consistently outperforms baselines, and in addition to performing well with the planner's approximation of pixel densities our approach also performs comparably for evaluation based on rendered views. Overall, the multi-round variant of the sequential planner we propose meets (within 1%) or exceeds the formation and assignment baselines in all scenarios we consider.