Multidimensional Skill Training for Human Users via Shared Control

This paper proposes a multidimensional skill-conditioned shared control framework for training human users in complex dynamic control tasks. In human training, novices often learn through repeated practice and expert guidance. However, this expert-novice paradigm may not scale well in real-world settings due to the scarcity of expert instructors. The proposed framework serves as an expert, automatically evaluating user skill levels and providing an appropriate amount of assistance by fusing the autonomous control input with the user's input. Complex tasks naturally involve multiple distinct control behaviors that may develop at different learning rates. To accommodate this, we propose a skill identification method that decomposes a given task into multiple distinct sub-skills and develop a discriminator-based strategy for evaluating user proficiency at the sub-skill level. The proposed skill-conditioned shared control scheme dynamically adjusts the contribution of the autonomous input for each sub-skill based on the estimated proficiency of a human user. It increases the control authority of the autonomous input when proficiency is low to prevent user frustration, and reduces it as proficiency improves to prevent over-reliance on autonomy. We demonstrate the proposed framework on a set of multi-rotor Unmanned Aerial Vehicle (UAV) trajectory tracking tasks using simulated expert and novice agents. The proposed framework is compared against no shared control and a baseline single-dimensional shared control. Results show that the proposed framework can adaptively provide targeted assistance based on individual sub-skill levels and reduce the trajectory tracking error compared to these baselines.