Data Driven Adaptive Autonomy

Abstract: Manual work in hazardous environments puts the operators at a great risk.Developing autonomous robots specialized for the task is one way to removethe need for human operators to be present at the scene makes but still lacksthe reasoning and expert knowledge of a human operator. Teleoperation isanother way to remove the operator from the dangerous environment but itas well has it’s drawbacks. The lack of direct feedback and latency reducesthe operator’s precise control and understanding of the task. Virtual Reality(VR) is a way to bridge this gap by improving the perceived feedback especiallyvisual feedback. Adaptive autonomy is the method that handles thecombination of autonomy and teleoperation by varying which side shouldget the most control at different moments in time in order to improve performancebased on different aspects dependent on the system. This thesisfocuses on the operator’s state of distraction to regulate the adaptive autonomysystem. This is done by exploring the possibilities of a data drivenapproach and proposing a suitable learning method that can classify whenthe operator is distracted in a VR teleoperation task with a simulated robotarm. There are two parts of this project, developing the data collection experimentand prototyping a LSTM learning method using the collected data.The learning method prototype showed promising results at predicting themoments when the operator was distracted.