What Is an End Effector in Robotics?

A robot policy does not command a hand in the abstract. It commands an end effector, often as a target pose or a change in pose, and the controller works out the joint motion. Human video shows a hand doing the task, but a hand has different degrees of freedom, grasp geometry and viewpoint from a gripper, so the footage does not say what a given robot should command. That gap is why published work adds a mapping step between human motion and an end-effector action.

Firsthand captures humans, not robots, and does not deliver robot actions, robot joint data, end-effector poses or teleoperation data. It captures the human-side signals nearest to end-effector motion: 21-joint 3D hand pose per hand with a visibility flag, wrist trajectories from a 6-DoF IMU on each wrist at 200 Hz, contact events and action segments, all on one PTP clock. These are collected fast, on demand and to a written skill spec. Any mapping to a robot is the buyer’s.

The end effector is the working end of a robot arm, the part that grasps, holds, presses or cuts. It can be a gripper with two fingers, a multi-fingered hand, a suction cup or a tool fixed to the arm. Everything before it, the links and joints, exists to place it where the task needs it.

The term separates what touches the object from what moves it. A policy that has to pick up a cup is judged by what the end effector does at the cup, not by the exact angle of each shoulder or elbow joint.

End-effector pose is where the end effector is and how it is oriented, relative to a chosen reference frame such as the robot’s base or a camera. Position is three numbers and orientation is three more, which is why it is often called a six-degree-of-freedom pose.

Many controllers accept a pose and solve for the joint angles that reach it, which is called inverse kinematics. That makes pose a convenient action: it stays meaningful across arms with different joint layouts. A pose can be absolute in a fixed frame or relative to the previous step, and the choice changes what the model has to learn.

Grippers close on an object, dexterous hands articulate several fingers for in-hand manipulation, suction cups hold smooth surfaces and tool attachments serve a single job. The table below lists them in general terms and records what this site states about each, which for most is nothing.

GR00T N1 (arXiv 2503.14734) describes a RoboCasa benchmark in which the state is “the position and rotation of both the end-effector and the robot base, as well as the gripper’s state,” and the action space is “the relative position and rotation of the end-effector along with the gripper state.”

EgoScale (arXiv 2602.16710) states that its “relative end-effector formulation removes dependence on absolute camera pose,” and controls its robot arms in a space where “actions specify incremental position and orientation changes, matching the wrist-pose representation used in human demonstrations.”

A human hand and a robot end effector differ in kinematics, degrees of freedom and sensor placement, so human motion does not map one-to-one onto robot commands. This is the embodiment gap, and a retargeting or alignment step is typically used to narrow it.

HumanScale (arXiv 2606.20521) describes how its egocentric data is labelled: “per-clip end-effector poses and gripper states are estimated from retargeted hand-pose signals as pseudo-action labels.” EgoVerse (arXiv 2604.07607) uses 3D hand pose and head pose as signals that “serve as proxies for human end-effector motion, enabling cotraining across embodiments.” These are each study’s own methods.

The site states human-side capture only. Per hand it records 21 joints in 3D, in metres in the head-camera frame, with a per-joint visibility flag. Each wrist carries a 6-DoF IMU at 200 Hz. Contact events mark when a hand took hold of or released an object, and action segments describe each sub-action with a verb and a noun, drawn from 97 verbs and 512 nouns. All streams share one PTP clock.

The site describes these as the fields an alignment or retargeting step would read. It does not describe them as end-effector poses.

Firsthand does not deliver robot actions, robot joint data, end-effector poses, gripper-state streams, teleoperation data or retargeted human-robot data. The site states no head or camera 6-DoF pose, so it gives no stated reference frame for converting hand motion into relative end-effector motion. It also states no accuracy figure for any mapping from its human signals to a robot.

Custom and on-demand collections are scoped by a written skill spec, so the tasks, settings and fields the mapping needs are written down before capture. Request a custom dataset and list the human-side signals your retargeting step reads. The robot side, including how to turn hand and wrist motion into end-effector actions, stays the buyer’s to plan.

What is an end effector in robotics?

The gripper, hand or tool at the end of a robot arm, the part that touches and acts on objects. The rest of the arm exists to position it.

What is end-effector pose?

The position and orientation of the end effector in a chosen reference frame. Policies often output it as a target pose or as a change in pose, and a controller solves for the joint motion.

Is a human hand an end effector?

No. A human hand and a robot end effector differ in kinematics, degrees of freedom and sensor placement, so human motion needs a retargeting or alignment step before it maps onto robot commands.

Does Firsthand deliver end-effector poses or robot actions?

No. Firsthand captures humans, not robots. It delivers human-side signals: 21-joint 3D hand pose, wrist IMU streams, contact events and action segments on one PTP clock. It does not deliver robot actions, joint data, end-effector poses or teleoperation data.

Which Firsthand signals are closest to end-effector motion?

Hand pose and wrist motion. Published studies use human hand pose as a proxy for end-effector motion, but each applies its own mapping, and the site states no mapping or accuracy figure for Firsthand data.

Can I order a custom human dataset for end-effector research?

Yes, for the human side. Custom and on-demand collections are scoped by a written skill spec. Request a custom dataset and list the hand, wrist and contact signals your mapping reads.