Intention Recognition in Training : Evaluating Smart Coaching in Machine Simulators

Sammanfattning

This master's thesis explores how intention recognition can be applied in training forest harvesting operators in simulators, focusing on the Komatsu Forest simulators produced by their partner Oryx Simulations. Intention recognition can be used to replicate the role of a human coach by identifying the operator's intentions, and provide them with suitable instructions and appropriate feedback. This study combines a state-of-the-art review on intention recognition with the design thinking process to create a human-controlled prototype of an intention recognition coach, referred to as a smart coach. The design thinking process has utilized user testing with the Wizard of Oz method, as well as an interview with an expert in the field of simulator training. During these user tests, the participants executed a training scenario in a Komatsu Forest simulator while receiving different instructions and feedback from a smart coach. The participants were informed that this smart coach was an automated intention recognition model, while the smart coach was entirely controlled by the researcher of this study. Each participant was then asked on their thoughts of the smart coach and how they would compare it to a human coach. Despite limitations concerning validity, testing conditions, and few participants the findings indicate that a smart coach could be helpful for less experienced operators and could improve user experience. The smart coach prototype was generally well-received by the seven participants that tested it. The results also indicate that a smart coach could only work as a supplemental teacher, and cannot replace a human coach entirely. This study offers valuable insights for future developments and iterations of a smart coach. The study finds that given the current state of intention recognition and especially the lack of data, the most viable solution to implement a smart coach is with a behavior tree, containing scripted instructions and feedback responses. This solution is presented as the Evidence-based Navigation Teacher that utilizes the extensive amount of operator metrics data that was discovered during this study.