LAUSR

We learn interpretable end-to-end controllers based on Neural Circuit Policies (NCPs) to enable goal reaching and dynamic obstacle avoidance in flight domains. In addition to being able to learn high-quality control, NCP networks are designed with a small number of neurons. This property allows for the learned policies to be interpreted at the neuron level and interrogated, leading to more robust understanding of why the artificial agents make the decisions that they do. We also demonstrate transfer of the learned policy to physical flight hardware by deploying a small NCP (200 KB of memory) capable of real-time inference on a Raspberry Pi Zero controlling a DJI Tello drone. Designing interpretable artificial agents is crucial for building trustworthy AIs, both as fully autonomous systems and also for parallel autonomy, where humans and AIs work on collaboratively solving problems in the same environment.