Robot End Effectors: Types, Specs & Where to Buy (2026 Guide)

What Is a Robot End Effector?

A robot end effector is the device attached to the terminal link of a robot arm — the part that touches and manipulates the world. The name comes from the functional role: it is the effector of the arm's motion, the component that actually effects change on objects in the environment. In industrial settings, end effectors are also called EOAT (end-of-arm tooling).

The robot arm itself provides positioning: it moves the end effector to the right location with the right orientation. The end effector provides actuation: it grabs, releases, pushes, cuts, welds, or otherwise interacts with the object. These are distinct engineering problems. A robot arm with a suboptimal end effector for the task will fail regardless of how good the arm kinematics are.

For robotics researchers and engineers building imitation learning systems or manipulation pipelines, the end effector choice is especially consequential. The training data format, the state representation, the reward function, and the task definition all depend on what kind of end effector you are using. Switching end effectors mid-project typically means collecting new data from scratch.

Types of End Effectors

1. Parallel Jaw Grippers

The most common end effector type in research and industry. Two parallel fingers close symmetrically around an object. Simple, reliable, and well-supported across all major robot arm platforms. The Robotiq 2F-85 is the industry standard: 85mm max opening, 2.5N to 220N force range, RS-485 and USB communication, force and position control. It is used in the vast majority of published imitation learning and pick-and-place research.

Parallel grippers excel at grasping objects with predictable geometry — cylinders, boxes, flat objects with consistent edges. They struggle with deformable objects (cloth, bags), objects with complex surface geometry, and tasks requiring fingertip-level precision. For most tabletop manipulation benchmarks, a parallel gripper is the right choice and any other option adds complexity without commensurate benefit.

Key variants: 2-finger (Robotiq 2F-85, OnRobot RG6), 3-finger (Robotiq 3F for cylindrical objects), wide-stroke (for large objects), miniature (for precision assembly).

2. Adaptive / Underactuated Grippers

Adaptive grippers use mechanically compliant fingers that conform to object shape during grasping. A single actuator drives multiple joints in each finger, with the joint motion governed by the contact mechanics rather than independent position control. This means the gripper can grasp irregular objects, spheres, and deformable items without needing precise pre-grasp positioning.

The Robotiq Hand-E is the most common industrial adaptive gripper, with parallel-jaw precision and adaptive capability in one unit. For research, adaptive grippers reduce the difficulty of robust grasp planning because contact tolerance is built in mechanically rather than requiring it from the policy.

3. Vacuum / Suction Cup Grippers

Suction grippers use vacuum pressure to grip flat, smooth, non-porous surfaces. They are the dominant end effector in logistics and e-commerce automation because boxes and flat-packed items are ideal suction targets. A well-designed vacuum system can achieve grasps in under 100ms and handle a wide range of object sizes without mechanical reconfiguration.

Suction grippers fail on porous materials (fabric, foam), curved surfaces where the cup cannot form a seal, and heavy objects that exceed the vacuum force. For research, they are primarily used in bin-picking and sorting applications where the object set is constrained to flat, smooth items.

4. Dexterous Multi-Finger Hands

Dexterous hands have 4-5 fingers with multiple actuated joints each — typically 12-20 DOF total. They can perform in-hand manipulation: rotating, repositioning, and reorienting objects while maintaining contact, without releasing and regrasping. This is the class of manipulation humans perform continuously with our hands and that simple grippers cannot approximate.

Dexterous hands are the most capable end effector type and the most complex. They require significantly more sophisticated control, produce higher-dimensional state and action spaces for learning algorithms, and are orders of magnitude more expensive than parallel grippers. The Allegro Hand (16 DOF, ~$15,000) and the Wuji Hand (11 DOF, ~$16,000) are the main research platforms. Both are available through SVRC.

Use cases: in-hand manipulation research, sign language synthesis, fine assembly tasks requiring finger-level contact control, teleoperation studies that require mapping human hand motion directly to robot hand motion.

5. Tactile Sensors and Fingertip-Integrated Sensors

Tactile sensors measure contact force distribution across a fingertip or finger surface, providing rich contact feedback that proprioceptive sensors (joint encoders) cannot capture. They range from simple force/torque sensors at the wrist (which provide 6-axis contact information for the whole end effector) to high-resolution tactile arrays embedded in individual fingertips.

For imitation learning, tactile feedback has been shown to substantially improve policy performance on contact-rich tasks — peg insertion, cable routing, screwing — where visual feedback alone is ambiguous at the moment of contact. The PaXini tactile sensor ($1,299) provides 3,750 sensing points per fingertip and is compatible with both Robotiq and custom gripper interfaces. SVRC also carries wrist-mounted force-torque sensors for simpler contact feedback applications.

6. Custom Tooling

Not all end effectors are grippers. Welding torches, spray nozzles, deburring tools, screwdrivers, and soldering irons are all end effectors — they are mounted on the arm and actuated as part of the task. In manufacturing automation, custom tooling often represents the majority of end effector deployments by volume. The OpenArm wrist kit provides a standard ISO 9283 flange interface for mounting custom tools.

How to Choose: Key Specifications

Choosing an end effector requires matching the mechanism to the task. The parameters below are the ones that actually determine whether an end effector works for a given application.

Payload Capacity

Payload is the maximum weight the end effector can lift and hold without mechanical failure or loss of position control. This is distinct from the arm payload — the arm rating typically includes the end effector weight, so the effective object payload is arm_payload minus end_effector_weight. For a ViperX 300 rated at 750g with a Robotiq 2F-85 weighing 900g, you are already over the arm's payload — you need to account for both numbers.

Stroke / Opening Width

For jaw grippers, stroke is the maximum jaw opening. The Robotiq 2F-85 opens to 85mm; the 2F-140 opens to 140mm. Your target objects need to fit within this opening when accounting for approach geometry — you often cannot approach at the object's widest point. For vacuum grippers, this parameter does not apply; capacity is determined by cup diameter and vacuum pressure.

Degrees of Freedom (DOF)

More DOF means more flexibility in grasp type and in-hand manipulation capability, but also higher control complexity, more failure modes, and higher cost. For most pick-and-place and structured manipulation tasks, 1-DOF (open/close) is sufficient and preferred for its simplicity. For dexterous manipulation research, 12-20 DOF is required. Most imitation learning benchmarks use 1-DOF grippers because the action space is already high-dimensional without adding end effector DOF.

Force Range and Controllability

The minimum controllable force matters as much as the maximum. A gripper that can exert 200N but has minimum force of 20N will crush eggs, circuit boards, and soft fruit. The Robotiq 2F-85 goes down to 2.5N of controllable force, which handles most objects. For delicate assembly tasks, look for grippers with force resolution below 1N. For high-throughput industrial tasks, you want maximum force and cycle time, with force range a secondary concern.

Communication Protocol

RS-485 (Modbus RTU) is the industrial standard. USB is simpler for research setups. EtherCAT is used for high-speed control loops. ROS2 driver availability matters significantly for research: the Robotiq 2F-85 has a well-maintained ROS2 driver; some less common grippers require custom driver development. Check that the gripper's control frequency matches your arm's control loop — mismatches create latency that shows up as jerky grasp execution in demonstrations.

Repeatability and Backlash

Repeatability specifies how consistently the gripper returns to a commanded position. For precision assembly tasks (inserting a USB plug, stacking thin discs), you need sub-millimeter repeatability. Most research grippers spec 0.02-0.1mm repeatability under ideal conditions; real-world repeatability is typically 2-5x worse due to wear and cable backlash. Grippers with direct-drive actuation (no gears) have lower backlash at the cost of higher motor current and cost.

Products at SVRC

All products below are available in SVRC's store or through our hardware catalog. Pricing is current as of May 2026. Contact us for volume pricing, leasing options, or integration support.

Robotiq 2F-85 — $5,825

The standard for manipulation research. 85mm stroke, 2.5-220N force range, 1-DOF, RS-485 + USB. The 2F-85 is used in more published imitation learning papers than any other gripper. It has a mature ROS2 driver, well-documented Modbus protocol, and extensive community support. If you are starting a new manipulation research project and do not have a specific reason to use something else, this is the correct default choice.

The 2F-85 integrates directly with ViperX, UR series, and Franka arms via ISO flanges. SVRC's store page at /store/product/robotiq-2f-85-gripper includes the flange compatibility matrix and links to our pre-configured ROS2 launch files for common arm combinations. See also the full hardware reference at /hardware/robotiq-2f-85/.

Wuji Hand — ~$16,000

An 11-DOF anthropomorphic hand designed for dexterous manipulation and teleoperation research. The Wuji Hand is used at SVRC for in-hand manipulation data collection and for mapping operator hand motion to robot hand motion during teleoperation. It supports force control on each joint and has a MuJoCo simulation model available for sim-to-real transfer research. ROS2 driver and Python SDK are maintained by Wuji Tech.

The Wuji Hand is the right choice when your task requires in-hand manipulation — rotating a pen, screwing a cap, handling deformable objects with fingertip-level control. For simpler grasping tasks, a parallel gripper is cheaper and easier to control. See the full specs at /hardware/wuji-hand/.

Allegro Hand — ~$15,000

A 16-DOF 4-finger dexterous hand from Wonik Robotics. 4 fingers, 4 DOF each, torque-controlled joints. The Allegro Hand is the most widely published dexterous hand platform in academic robotics — it appears in papers on dexterous in-hand manipulation, reinforcement learning for hands, and tactile-based grasping. The large research community means there are open-source RL environments, pre-trained policies, and sim models freely available.

The Allegro's primary advantage over the Wuji Hand is community breadth: more published baselines, more open-source code, and more direct comparisons with other work. The Wuji Hand has a more anthropomorphic finger layout and stronger finger force. For teams planning to publish against existing Allegro baselines, the Allegro is the right choice. For teams building teleoperation systems that map from a human glove, the Wuji Hand's layout may be more compatible. See the hardware reference at /hardware/allegro-hand/.

PaXini Tactile Sensor — $1,299

A fingertip-mounted tactile sensor with 3,750 sensing points per fingertip, 500Hz sampling rate, and USB interface. The PaXini sensor fits on standard Robotiq fingertip mounts and provides a dense pressure map that captures contact shape, force magnitude, and slip detection. For contact-rich manipulation tasks — peg insertion, screw driving, pen handling — tactile feedback dramatically reduces the number of demonstrations needed to train a successful policy.

The PaXini sensor is one of the most cost-effective ways to add tactile feedback to an existing gripper setup without changing the end effector itself. SVRC includes PaXini data channels in our standard data collection configurations. See specs at /hardware/paxini-tactile/.

OpenArm Wrist Kit

A standardized wrist interface module for the OpenArm 101 that includes a 6-axis force-torque sensor, ISO 9283 tool flange, and breakout electronics for custom tooling. The wrist kit is designed for teams building custom end effectors or integrating commercial grippers with the OpenArm platform. It is included with the SVRC DK1 data collection kit and is available separately for OpenArm upgrades. See details at /hardware/openarm-101/.

Comparison Table: Major Research End Effectors

Use Cases: Pick-and-Place, Teleoperation, and Imitation Learning

Pick-and-Place Automation

For structured pick-and-place — known objects, fixed orientations, predictable bin locations — a 1-DOF parallel gripper is almost always correct. The Robotiq 2F-85 handles the majority of industrial pick-and-place objects up to 85mm in diameter. Vacuum grippers are preferred when the object set is flat, smooth, and varied in shape (e-commerce fulfillment). Adaptive grippers like the Robotiq Hand-E are used when objects vary in shape but are too porous or curved for suction.

The key pick-and-place end effector error is over-specifying: buying a dexterous hand for a task that a parallel gripper handles perfectly is expensive and introduces control complexity that does not improve outcome. Start with the simplest end effector that physically handles your objects before considering more complex options.

Teleoperation and Data Collection

For human teleoperation data collection, the end effector must map naturally to the operator's control input. With a parallel gripper, the operator's trigger or button controls open/close. With a dexterous hand, the operator typically wears a glove (like the Wuji Glove) and the hand mirrors finger positions in real time. The complexity of the control interface scales with the end effector DOF.

For most imitation learning data collection programs, the 2F-85 is the default because the control interface is simple (one axis), operators require minimal training, and demonstrations are easier to collect consistently. Dexterous hand data collection requires trained operators with 20-40 hours of practice to produce quality demonstrations. SVRC's managed data collection service has experienced operators for both parallel and dexterous end effectors. See the Wuji Hand teleoperation guide for dexterous hand data collection specifics.

Imitation Learning and Policy Training

The end effector you choose shapes everything about your learning pipeline. A 1-DOF parallel gripper adds one dimension (gripper position) to the action space. A 16-DOF dexterous hand adds 16 dimensions — and each of those dimensions needs to be covered adequately in the demonstration data. Data requirements scale roughly linearly with end effector DOF for structured tasks and superlinearly for contact-rich tasks where the fingers interact simultaneously.

For researchers starting with imitation learning, the recommended path is: start with a parallel gripper and get your first policy working end-to-end (collecting demos, training, evaluating) before adding end effector complexity. Once you understand your data collection throughput and policy training pipeline, you can evaluate whether dexterous end effectors are worth the investment for your specific task.

Diffusion Policy and ACT both handle 1-DOF gripper action spaces trivially. VLAs (OpenVLA, pi0) were trained on datasets that include Allegro Hand and other dexterous end effectors, which means there is more pretrained knowledge in those models for dexterous manipulation — but getting usable fine-tuning data for dexterous tasks is significantly harder. See our ACT vs Diffusion Policy guide for policy architecture considerations.

Frequently Asked Questions

What is a robot end effector?

A robot end effector is the device mounted at the end of a robot arm that interacts with the environment. It is the mechanical equivalent of a hand — it grasps, moves, or manipulates objects. Common types include parallel jaw grippers, vacuum cups, dexterous multi-finger hands, and custom tooling like welding torches or spray nozzles.

What is the difference between a gripper and an end effector?

All grippers are end effectors, but not all end effectors are grippers. A gripper is a specific type of end effector that grasps and holds objects using mechanical fingers or suction. Other end effectors include welding tools, paint sprayers, force-torque sensors, and dexterous hands. In robotics research, the terms are often used interchangeably when referring to manipulation tasks.

How do I choose a robot gripper for imitation learning?

For imitation learning and teleoperation, the key factors are: (1) DOF — more degrees of freedom enable more complex grasps but require more complex control; (2) compatibility with your arm's flange and protocol (RS-485, USB, ROS); (3) whether you need tactile feedback for contact-rich tasks; (4) payload and stroke range for your target objects. For most tabletop manipulation research, a 2-finger parallel gripper like the Robotiq 2F-85 is the standard starting point.

What end effectors does SVRC carry?

SVRC carries the Robotiq 2F-85 ($5,825), Wuji Hand (~$16,000), Allegro Hand (~$15,000), PaXini tactile sensor ($1,299), and OpenArm wrist kit. All are available in our store or through the hardware catalog. We can also help you evaluate which end effector is right for your task — contact us to discuss your application.