Towards Latency Efficient DRL Inference: Improving UAV Obstacle Avoidance at the Edge Through Model Compression

It is vital that autonomous Unmanned Aerial Vehicles (UAVs) are able to avoid obstacles effectively. When avoiding such obstacles it is important that movement decision are made fast (i.e. inference latency is low) so that crashes are avoided. When deep reinforcement learning (DRL) is being leveraged as the method of obstacle avoidance one way of reducing this inference latency is to deploy the DRL model at the edge (e.g., on-UAV). However, even if the DRL model is small enough to be deployed on-UAV, the inference latency can be too high. There is a lack of research that examines reducing DRL inference time of UAVs when avoiding obstacles. Thus, this paper examines various model compression techniques to improve the inference speed of such DRL models deployed at the edge. We propose an approach that combines these model compression techniques and apply it to a well performing Dueling Double Deep Q-Network (D3QN) baseline model. On the Nvidia Jetson Orin Nano and Nvidia Jetson Nano edge devices we show that, relative to our baseline model, this combined model compression approach reduces inference latency by 38.61% and 53.18% while only reducing the success rate by 2.34% and 5% respectively. All our code is open-sourced.