Human-robot planar co-manipulation of extended objects: data-driven models and control from human-human dyads

Area of Science:

• Robotics
• Human-Computer Interaction
• Control Theory

Background:

• Human-robot collaboration in manipulating large objects is challenging due to motion ambiguity.
• Existing methods struggle to accurately infer human intent during physical co-manipulation tasks.

Purpose of the Study:

• To leverage human-human dyad data for predicting motion intent in human-robot co-manipulation.
• To develop and evaluate novel controllers for robot-assisted manipulation tasks with ambiguous motion goals.

Main Methods:

• Analyzing torque patterns in human-human object manipulation to identify intent signals.
• Developing a deep neural network using human-human motion data to predict future object trajectories.
• Integrating force and motion data for real-time robot control in human-robot dyads.
• Evaluating controller performance in three-degree-of-freedom planar motion tasks.

Main Results:

• Distinct torque triggers were identified in human-human data for lateral movements, indicating intent.
• A deep neural network effectively predicted future object motion based on past data from human-human trials.
• The developed controllers demonstrated effective performance in ambiguous human-robot co-manipulation scenarios.

Conclusions:

• Human-human interaction data provides valuable insights for inferring motion intent in human-robot collaboration.
• Data-driven approaches, including deep learning, can significantly improve robot control for physical co-manipulation.
• The proposed methods enhance the intuitiveness and efficiency of human-robot physical interaction.