Versatile kinematics-based constraint identification applied to robot task reproduction.

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Identifying kinematic constraints is crucial for robot task execution.
• Existing methods often require specific prior information or measurements, limiting their applicability.
• A versatile, information-agnostic approach is needed for real-world robotic systems.

Purpose of the Study:

• To propose a versatile kinematics-only method for identifying robot-environment constraints.
• To enable constraint identification without prior knowledge of constraint models, geometry, or force measurements.
• To demonstrate the method's effectiveness in simulation and real-world robot experiments.

Main Methods:

• Utilizes constraint reference frames attached to robot or ground bodies.
• Identifies constraints by minimizing velocity norms in the Cartesian components within these frames.
• Applies a kinematics-only approach, relying solely on measured kinematics.

Main Results:

• Successfully identified the geometry of twelve diverse constraints in simulation, including articulated, polyhedral, and contour-following contacts.
• Demonstrated that accuracy decreases linearly with increasing sensor noise.
• Achieved comparable task reproduction performance in robot experiments to existing literature methods.

Conclusions:

• The proposed kinematics-only method offers a versatile solution for identifying robot-environment constraints.
• The approach is applicable to various robots and environments lacking prior constraint information, suitable for everyday robotic applications.
• This method enhances autonomous task execution by enabling robots to learn and adapt to environmental constraints.