Multifingered Robot Hand Compliant Manipulation Based on Vision-Based Demonstration and Adaptive Force Control.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Dexterous manipulation with multifingered robot hands remains a significant challenge.
• Achieving compliant behaviors in robotic hands is crucial for safe and effective interaction.

Purpose of the Study:

• To develop a human-in-the-loop learning-control approach for acquiring compliant grasping and manipulation skills in multifingered robot hands.
• To enable robot hands to exhibit human-like compliant behaviors through learned control strategies.

Main Methods:

• A markerless vision-based teleoperation system using depth images of a human hand for task demonstration.
• An end-to-end neural network (TeachNet) trained to map human hand pose to robot hand joint angles in real-time.
• An adaptive force control strategy based on biomimetic human motor learning principles to predict force commands and adapt impedance/feedforward profiles.

Main Results:

• The proposed approach successfully trained a robot hand (Shadow Hand) to perform compliant grasping and manipulation.
• The system demonstrated more reliable performance compared to traditional position control modes.
• The adaptive force control strategy enabled compliant interaction with the environment.

Conclusions:

• The human-in-the-loop learning-control approach is effective for imparting compliant manipulation skills to multifingered robot hands.
• Biomimetic force control strategies can enhance the human-like compliance of robotic systems.
• This method offers a promising alternative to position control for achieving compliant robotic behaviors.