Kinesthetic Teaching

• Also called: lead-through programming, hand guiding, physical demonstration
• Category: Learning & autonomy › Learning from demonstration
• Tags: imitation-learning, demonstrations, human-robot-interaction, cobots, data-collection

Teaching a robot by physically guiding its arm through a task while it records its own joint or end-effector motion, giving demonstrations directly in the robot's own body.

Definition

Kinesthetic teaching is a way of giving demonstrations in which a person holds the robot and moves it through the task. The robot's controller makes the arm easy to move, and the robot records its own joint positions, end-effector poses and sometimes forces as the demonstration. The recorded data can be replayed as a fixed program or used as training data for imitation learning. Kinesthetic teaching is therefore a way of collecting demonstrations, while imitation learning is the family of methods that learn a policy from them, whichever way they were collected. In industrial robotics the same idea is called lead-through programming or hand guiding.

The main advantage is that demonstrations come out directly in the robot's own joint space. Teleoperation and video demonstrations require mapping human motion onto a different body, the correspondence problem, and teleoperation interfaces can be hard to use for fine or contact-rich motion. With kinesthetic teaching, every recorded motion is already kinematically feasible for the robot, and the person feels the contact with the environment through the arm. To make this possible the arm must be easy to move by hand: either it is backdrivable and runs a gravity-compensation or low-impedance controller, or it uses joint torque or wrist force sensing with admittance control to follow the forces the person applies.

Kinesthetic teaching also has limits. The person's hands and body are in the scene, so images recorded during the demonstration contain the person and the forces recorded include the person's push, not only the task forces. A common fix is to record the trajectory kinesthetically and then replay it autonomously to capture clean camera and force data. Guiding a heavy or many-jointed arm is tiring, two-arm tasks need two people or careful setups, and fast or dynamic motions are hard to demonstrate by hand. Demonstrations also capture the teacher's hesitations, so trajectories are usually smoothed or segmented before use.

Why it matters

Kinesthetic teaching is one of the simplest ways to get feasible, task-relevant demonstrations onto a real robot, with no retargeting and little extra hardware. It is widely used both for programming collaborative robots on the factory floor and for collecting data for imitation learning research.

Key applications

• Programming collaborative robots by hand guiding
• Collecting demonstrations for imitation learning and behavior cloning
• Teaching contact-rich tasks such as wiping, insertion and scooping
• Learning movement primitives from a few demonstrations
• Teaching waypoints for welding, painting and assembly paths

In practice

A research team collects demonstrations of scooping food from a bowl with a spoon attached to a collaborative arm. In gravity-compensation mode, a person grasps the arm near the wrist and performs each scoop while the robot records end-effector poses at a fixed rate. Because the person's hand blocks the wrist camera and adds force, each recorded trajectory is then replayed by the robot alone, and the camera images and force readings from the replay are paired with the trajectory as training data. Demonstrations that hit the bowl rim or spill food are discarded, and the rest are used to train a visuomotor policy.

Common misconceptions

• Kinesthetic teaching is a kind of teleoperation — in teleoperation the person moves a separate interface and the robot follows remotely, while in kinesthetic teaching the person physically moves the robot itself
• Any robot arm can be taught kinesthetically — the arm needs backdrivable joints or torque or force sensing with a suitable controller; a stiff position-controlled industrial arm cannot simply be pushed by hand
• Data recorded during the demonstration can be used as is — the person appears in camera images and their applied force mixes with task forces, so many pipelines replay the trajectory to record clean sensor data

Who should know this

• Robot learning researchers collecting demonstrations
• Integrators programming collaborative robots
• Controls engineers implementing gravity compensation and admittance control
• Assistive and service robotics developers

Related technologies

• Gravity Compensation
• Admittance Control
• Dynamic Movement Primitives
• Joint Torque Sensors
• Collaborative Robot Arms

Related terms

• Imitation Learning
• Teleoperation
• Behavior Cloning
• Backdrivability
• Impedance Control
• Cobot

Related projects

• LeRobot: HuggingFace ML library for real-world robotics in PyTorch
• OpenPI: Physical Intelligence VLA foundation models for general robot control
• Mobile ALOHA: Bimanual mobile manipulation with 4 robot arms on a wheeled base for teleoperation
• OpenArm: 7DOF humanoid arm for physical AI research in contact-rich environments
• Koch v1.1 Low-Cost Robot Arm: Open-source 6-DOF robot arm for affordable teleoperation and imitation learning research.
• AgiBot World: Million-trajectory real-robot manipulation dataset from a fleet of 100+ humanoid robots