Kinesthetic Teaching

The defining characteristic of kinesthetic teaching is the direct physical manipulation of the robot's end-effector or arm by a human operator. This bypasses traditional programming interfaces, allowing the teacher to impart motion skills through haptic demonstration. The robot operates in a zero-force gravity-compensated mode or a back-drivable compliant state, making it easy to move.

• Example: An operator grasps a collaborative robot's gripper and manually guides it through the precise motions needed to insert a peg into a hole.
• Key Benefit: Enables teaching of tasks that are difficult to describe symbolically but are intuitive to perform physically.

During physical guidance, the robot's control system records the demonstrated trajectory at a high frequency. This includes the sequence of joint positions, end-effector poses (position and orientation), and sometimes velocities and forces. This recorded path becomes the reference trajectory for autonomous execution.

• Data Captured: Timestamped joint angles, Cartesian poses, and optionally wrench (force/torque) data.
• Execution: The robot later replays this trajectory using a position or impedance controller. Advanced systems may smooth the raw data or use it to seed a dynamic movement primitive for generalization.

For effective kinesthetic teaching, the robot must exhibit mechanical and control compliance. This is achieved through:

• Hardware Back-drivability: Low-friction, low-gear-ratio actuators that can be easily moved when power is off or motors are in a compliant mode.
• Software Compliance: Control modes like gravity compensation, impedance control, or admittance control that make the robot feel light and responsive to human force.

This compliance is essential for a natural teaching experience and is a hallmark of collaborative robots (cobots) designed for this interaction.

Kinesthetic teaching is a primary data collection method for Imitation Learning. The recorded demonstrations are not just for playback; they serve as training data for more robust policies.

• Behavioral Cloning: The recorded state-action pairs are used to train a supervised learning policy to map states to actions.
• Inverse Reinforcement Learning: The demonstrations are analyzed to infer the reward function the teacher was optimizing, which can then be used for reinforcement learning.
• Dynamic Movement Primitives (DMPs): The trajectory is encoded into a parameterized model that can be adjusted for new start/goal positions.

It is crucial to distinguish kinesthetic teaching from related methods:

• Vs. Teleoperation: In teleoperation, the human uses a separate input device (joystick, haptic master). Kinesthetic teaching involves direct contact with the robot itself.
• Vs. Waypoint/Lead-Through Teaching: While similar, waypoint teaching often involves moving the robot to discrete points (using a teach pendant or hand-guiding) and saving those positions. Kinesthetic teaching emphasizes continuous path recording of the entire motion.

Kinesthetic teaching provides a high-dimensional, continuous demonstration ideal for learning smooth, dynamic skills.

This method is pivotal in domains requiring rapid, flexible task programming without code.

• Small-Batch Manufacturing: Quickly teaching a robot a new assembly or finishing task for a custom product run.
• Polishing & Deburring: Teaching the exact force and motion profile needed for surface treatment.
• Assistive Robotics: Enabling a caregiver or user to physically guide a robotic arm to perform a personal care task, which it can then repeat autonomously.
• Surgical Robotics: Surgeons may use kinesthetic guidance to define safe boundaries or optimal paths for robotic tools.

The technique bridges the gap between human motor skill and robotic precision.

Learning from Demonstration (LfD), also known as Programming by Demonstration or Imitation Learning, is the overarching machine learning paradigm where a robot acquires a skill by observing examples provided by a teacher. Kinesthetic teaching is a primary data collection method within LfD, where the demonstration is provided through direct physical guidance. Other LfD methods include:

• Teleoperation: Using a remote control interface.
• Shadowing: The robot observes a human performing the task via vision sensors.
• Behavioral Cloning: The robot learns a direct mapping from states to actions by mimicking the demonstrator's policy.

Hand Guiding is the specific collaborative robot operation mode that enables kinesthetic teaching. In this mode, the robot's joints enter a zero-gravity or compliant state, allowing a human operator to grasp the end-effector or arm and physically move it through a desired path. The robot's internal sensors (encoders, torque sensors) record this trajectory in joint space or Cartesian space. This mode is a foundational feature of collaborative robots (cobots) and requires force/torque sensing and backdrivable actuators to function smoothly and intuitively.

Physical Human-Robot Interaction (pHRI) is the subfield focused on direct, intentional contact and force exchange between a human and a robot. Kinesthetic teaching is a quintessential pHRI application, as it relies on safe and intuitive physical contact. pHRI research addresses:

• Contact dynamics and force control.
• Safety standards like ISO/TS 15066.
• Hardware design for inherent safety (e.g., rounded edges, compliant joints).
• Control strategies for admittance and impedance control that make the robot compliant to human guidance.

A Collaborative Robot (Cobot) is a robot designed from the ground up to operate safely alongside humans in a shared workspace, making it the ideal hardware platform for kinesthetic teaching. Key features that enable this include:

• Power and Force Limiting (PFL) joints to prevent injury on contact.
• Inherently safe design with no pinch points and soft surfaces.
• Integrated force/torque sensing in the joints or wrist.
• Hand guiding functionality as a core operational mode.
• Safety-rated monitored stop to pause operation when a human gets too close.

Shared Autonomy is a control paradigm that dynamically blends human input with robot autonomy. While kinesthetic teaching is used for initial programming, shared autonomy is often employed during task execution. The system interprets the human's physical guidance not as a strict trajectory to replay, but as an intent signal. The robot's autonomous controllers then assist by:

• Filtering human jitter and smoothing the path.
• Adding constraints (e.g., keeping a drill perpendicular to a surface).
• Handling low-level stability while the human provides high-level directional input.
• This creates a synergistic partnership where the human and robot co-manipulate an object or tool.

Kinesthetic teaching provides the demonstration data for Imitation Learning (IL), which is often contrasted with Reinforcement Learning (RL) for robot skill acquisition.

• Imitation Learning (via Kinesthetic Teaching): The robot learns a policy from expert demonstrations. It's sample-efficient for learning specific tasks but is limited by the quality and coverage of the human teacher's demonstrations.
• Reinforcement Learning: The robot learns through trial-and-error, optimizing a reward function. It can discover novel, optimal strategies but is often extremely sample-inefficient and unsafe for physical robots. Hybrid approaches are common, where a policy initialized via kinesthetic teaching is then refined and made more robust using RL in simulation (sim-to-real).