Kinesthetic Teaching
How it works
The robot is put into a compliant/backdrivable mode with gravity compensation so a human can move its links with little effort. The operator physically guides the end-effector/arm through the task (e.g. grasp, transfer, insertion). During the motion the joint encoders sample the trajectory — positions, and often velocities and torques — at a fixed rate; some systems let the teacher record only keyframes instead of a continuous trajectory. The recorded demonstrations (often several repetitions) are encoded and generalized into a policy, e.g. via Dynamic Movement Primitives or GMM/GMR, enabling the motion to be reproduced and adapted to new initial conditions.
Problem solved
Programming a robot by traditional means (hand-coding trajectories, a teach pendant) is time-consuming and requires expertise. Demonstration by observing a human, in turn, requires solving the correspondence problem — mapping human motion onto the robot’s different kinematics. Kinesthetic demonstration addresses both: a non-expert can teach the robot intuitively, and the data is captured directly in the robot’s own joint space.
Components
Control that puts the robot into a backdrivable state in which a human can move its links with little effort while the arm’s weight is compensated.
Sampling joint encoder readings (positions, often velocities and torques) during the demonstration; optionally recording only keyframes.
Turning recorded trajectories into a control policy, e.g. via Dynamic Movement Primitives, Gaussian Mixture Models (GMM/GMR), or keyframe-based learning.
Implementation
Guiding many joints at once by hand is difficult; the method is poorly suited to fast/dynamic motions and to robots with many degrees of freedom (e.g. multi-fingered hands, whole-body motions).
Human demonstrations vary across repetitions and can be suboptimal, which lowers the quality of the learned policy.
Evolution
The Argall et al. survey organizes demonstration-provision methods (e.g. teleoperation, shadowing, guidance), situating kinesthetic demonstration within learning from demonstration.