Assistive Teleoperation for Manipulation Tasks
How should a human user and a robot collaborate during teleoperation? The user understands the full semantics of the task: they know, for example, what the robot should search for in a cupboard, or that it should be more careful when moving near a glass of water than near a box of tissues. Since the robot might not have this knowledge, allowing it to operate fully autonomously may be risky; its model is incomplete and its policy might be wrong. On the other hand, teleoperating the robot through every motion is slow and tiresome, especially on difficult tasks. Between these two extremes lies a spectrum, from almost no assistance at all (very timid) to full autonomy (very aggressive). So what is the appropriate level of assistance? And how do factors like task difficulty and policy correctness affect this decision?