End Effector

Conceptual

An end effector is the device mounted at the tip of a robot arm that physically interacts with the environment. It is the last element in the kinematic chain — the part that actually does the work. Examples include grippers, welding torches, suction cups, spray nozzles, and sensors.

“The arm gets the end effector to a position. The end effector does the work.”

The term covers anything that terminates the arm: a two-fingered gripper on a pick-and-place robot, a MIG torch on a welding robot, a vacuum cup on a palletizer, or a force-torque sensor used to detect contact during assembly. The choice of end effector is a design decision as consequential as the choice of robot arm itself.

Tip

End effectors are sometimes called tools or tooling, especially in manufacturing contexts. The point where the tool attaches to the robot arm is called the flange.

Types of End Effectors

Type	Actuation	Primary Application
Parallel gripper	Two fingers, single actuator	General-purpose manipulation
Multi-finger hand	3–5 actuated fingers	Dexterous manipulation
Suction cup	Vacuum pump	Flat objects, boxes, warehousing
Magnetic	Electromagnet	Metal parts, automotive
Tool changer	Mounts interchangeable tools	Flexible manufacturing
Force-torque sensor	Measures contact forces	Assembly, grinding, insertion
Camera	Mounted for visual inspection	Quality control, bin picking

Note

A tool changer is itself an end effector — one that can be swapped mid-task for a different tool. Industrial cells commonly use a tool changer rack so a single robot arm can weld, drill, and inspect in sequence without human intervention.

The Gripper Coordinate Frame

Motion planning and inverse kinematics do not reason about the robot flange — they reason about the Tool Center Point (TCP), a coordinate frame you define that represents the functional tip of the end effector.

For a parallel gripper the TCP is typically placed between the fingertips at the midpoint of the grasp. For a welding torch it is placed at the tip of the electrode. You tell the robot “move to this pose” and the robot computes the joint angles required to put the TCP there.

The full coordinate frame hierarchy is:

World frame (fixed)
  └─ Base frame (robot base)
       └─ Joint frames (one per joint)
            └─ Flange frame (last joint)
                 └─ Tool frame = End Effector position

Each frame is related to its parent by a rigid transformation (rotation + translation). The chain of transformations from the world frame to the tool frame is exactly what forward kinematics computes.

“When you command a robot to ‘move to (0.5, 0.3, 0.2)’, you’re commanding the tool frame to that position.”

Caution

If the TCP is defined incorrectly — offset or rotated from the true grasp point — every pick will miss by a consistent amount. TCP calibration is a required step when mounting a new gripper.

Relevance to Robot Learning

In imitation learning datasets, observation.state contains joint positions for each joint in the arm. The end effector contributes an additional channel that does not map to a joint angle in the usual sense.

For a simple parallel gripper this is typically a single float:

0.0  →  fully closed (grasping)
1.0  →  fully open

This float appears alongside the arm joint positions in the state vector and in the action vector (the policy must learn when to open and close the gripper, not just where to move the arm).

The gripper cycle — open, move to object, close around object, lift — is the critical success criterion for a pick task. An imitation learning dataset that contains demonstrations where the gripper never closes on the object (incomplete demonstrations) will train a policy that reaches toward objects but never grasps them. Validating that every episode contains at least one full open-to-close cycle is a mandatory data quality check before training.

Tip

In the SO-101 robot setup used in this platform, the gripper state appears as the gripper key in observation.state. Its value is in motor position units (radians), not normalized 0–1, but it plays the same role: a scalar channel encoding how open or closed the end effector is.

Related Terms

Kinematics How robots move through space — forward and inverse kinematics

Joints and Links The rigid bodies and articulated connections that make up a robot arm

Teleoperation Human-guided robot control used to collect demonstration data

Degrees of Freedom How many independent ways a body can move

Sources

• Siciliano et al., Robotics: Modelling, Planning and Control — Standard graduate textbook; Chapter 2 covers the kinematic chain and tool frames
• ROS 2 TF2 — Working with Sensor Data — Practical guide to coordinate frame transforms including the tool frame