CAREER: Towards Environment-Aware Adaptive Safety for Learning-Enabled Multiagent Systems with Application to Target Drone Capturing

There are increasing threats from unauthorized and malicious drones with research and industry communities looking for solutions. However, current anti-drone techniques are often prone to failure, not cost effective, or could affect legitimate nearby aircraft. This proposal develops a Multi-UAV Drone Catch Net (MUCH-Net) system that uses a team of low-cost autonomous unmanned aerial vehicles to collaboratively tether a catch net to capture the target drone.

This new system puts a high demand on the safe operations of the aerial vehicle team using a learning-based cooperative formation architecture design with environment-aware adaptive safety constraints. The broader impacts of the project include (a) a concept design contest for high-school pre-engineering program students, (b) a robot capture game competition (c) a robotic program for K-12 teachers, (d) promotion of exploratory learning assignments in undergraduate teaching, (e) a new graduate course on safety-critical intelligent multiagent systems, (f) and collaborations with industry partners to facilitate research development, verification, assessment, and technology transfer.

This CAREER project aims to make fundamental contributions to theories of learning-based cooperative control with new environment-aware adaptive safety analysis. Major technical challenges include: (a) due to the complex operating environment, the safety considerations are environment-aware and adaptive; and (b) for multiagent systems, the multiple safety requirements considered can be conflicting with each other or with the initial system state.

Existing safety-critical control algorithms for multiagent systems only address constant or time-varying safety sets, which cannot dynamically adapt to the environment, and cannot address safety conflicts. The research investigates learning-based cooperative control architectures to address environment-aware adaptive safety requirements for multiagent systems.

An integrated barrier function structure that integrates a cooperative dynamic deep neural network to learn the dynamics of a multi-dimension environment parameter and unknown target velocity is proposed. Safety conflicts are addressed by integrating initial state and virtual barriers into the integrated barrier functions, with indicator functions incorporated to modify the less critical (“soft”) safety sets, in order to guarantee the more critical (“hard”) safety requirements.

The proposed architectures are widely applicable to many applications with multiagent systems operating in complex environments. This project is jointly funded by the Electrical, Communications and Cyber Systems Division (ECCS) and the Established Program to Stimulate Competitive Research (EPSCoR).

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.