Shared Autonomy via Hindsight Optimization.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Shared autonomy combines user input and robot control for task achievement.
• Robots often lack prior knowledge of user goals, requiring prediction and assistance.
• This necessitates advanced methods for seamless human-robot interaction.

Purpose of the Study:

• To develop and evaluate a novel shared autonomy framework that addresses uncertainty in user goals.
• To improve robot assistance by accurately predicting and adapting to user intentions.
• To enhance task efficiency and user experience in human-robot collaboration.

Main Methods:

• Formulated shared autonomy as a Partially Observable Markov Decision Process (POMDP) with goal uncertainty.
• Utilized maximum entropy inverse optimal control to infer user goal distributions from input history.
• Employed hindsight optimization for approximating intractable POMDP solutions.

Main Results:

• The proposed method enabled users to accomplish tasks significantly faster compared to a predict-then-blend approach.
• Users required less input when utilizing the new shared autonomy system.
• User feedback indicated a perceived tradeoff between control authority and task completion speed.

Conclusions:

• The developed shared autonomy method effectively enhances task efficiency in human-robot interaction.
• While improving performance, user perception of control remains a key consideration in shared autonomy design.
• Future work should explore balancing user control and system efficiency for optimal collaboration.