A First CNN-based Approach Towards Autonomous Flight for Object Lifting

Abstract

Cable-suspended load transportation with Micro Air Vehicles (MAV) is a well-studied topic as it reduces mechanical complexity, the weight of the system, and energy consumption. However, it is always taken for granted that the load is already attached tocable. In this work, we present a methodology to autonomously lift a cable-suspended load with a MAV using a Deep-Learning based Object Detector as the perception system, whose detections are used by a PID controller and a state machine to perform the lifting procedure. We report an autonomous lifting success rate of 40%, an encouraging result considering that we carry out this task in a realistic environment, not in simulation. The Object Detector model has been tailored to detect the 2D position and 3D orientation of a bucket-shaped load and trained with a fully synthetic dataset. However, the model is successfully used in the real world. The control system deals with the oscillatory behavior of the cable and ground effects using low-level controllers. Future work includes improvements to the perception system to also detect a hook-shaped grasper.