Data, Hardware & Evaluation
Teleoperation Rigs
ALOHA, GELLO, UMI, and VR teleop: cost, data quality, throughput, and the embodiment gap.
Last reviewed 2026-08-09
Nearly every demonstration in a robot learning dataset was produced by a human moving hardware. The human drives the robot itself, moves a scaled twin of it, holds a gripper proxy in hand, or works inside a headset while the robot mirrors them. Four rig families cover most of that collection in 2026: ALOHA-class bimanual workstations, GELLO kinematically matched exoskeletons, UMI handheld grippers, and VR teleoperation. Each family is a different bet on four axes: what the rig costs, how faithfully its demonstrations map onto robot execution, how cheaply its demonstration volume scales, and how large the gap is between the recorded motion and the robot's. Prices span two orders of magnitude, from GELLO's sub-$300 parts list to a $32,000 ALOHA 2 configuration Wu 2023 alpibrusl 2026.
The comparison matrix
4 of 4 rigs
| Sources | |||||
|---|---|---|---|---|---|
GELLOKinematically matched exoskeletonGELLO for Franka, UR5, and xArm | $300Parts BOM under $300; excludes the target robot arm | highA scaled kinematic twin of the target arm; joint readings map one-to-one onto robot commands | mediumUnder $300 and about 30 minutes to assemble, but each target arm model needs its own build | lowThe replica shares the arm joint structure, so the operator feels the arm constraints directly | GELLO paperProject site |
UMIHandheld gripper, no robot at collectionUMI gripper with GoPro wrist camera | $371$73 printed gripper + $298 GoPro and accessories, per gripper | medium155-degree fisheye wrist camera with SLAM-recovered gripper pose; no force channel | highAbout 30 seconds per demonstration; over 3x faster than spacemouse teleoperation on the paper benchmark | mediumA handheld gripper stands in for the robot gripper; latency matching and relative-trajectory actions close the gap at deployment | UMI paperProject site |
ALOHA-class workstationBimanual leader-follower workstationALOHA 2 (Stanford), Trossen AI Stationary and Mobile AI | $17,000ALOHA 2 runs $17,000-$32,000; Trossen AI bimanual rigs $15,995-$22,995 | highLeader and follower arms share kinematics; demonstrations land directly in the robot joint space at 500 Hz | lowOne operator per fixed workstation; every additional collector costs another full rig | lowThe operator drives a kinematically identical arm, so recorded motion is the robot motion | ACT paperTrossen AI |
VR teleoperationHeadset plus retargeted human motionDROID Quest 2 rig, Open-TeleVision, Bunny-VisionPro | n/a | mediumImmersive stereo view and mirrored motion; actions are retargeted from the operator, not recorded from robot joints | highDROID: 50 operators across 13 institutions collected 76,000 trajectories (350 hours) in 12 months | highHuman hand and controller poses must be remapped through IK onto robot joints; calibration and avoidance layers exist because the mapping drifts | DROIDOpen-TeleVisionBunny-VisionPro |
Three reading rules for the matrix. First, ratings run low, medium, high, and the note in each cell states what the rating means for that rig and dimension, because "high throughput" is good news and "high embodiment gap" is not. Second, the highlight buttons swap the legend and print the per-rig detail for one dimension at a time. Third, every column sorts, cells with no published value are marked n/a, and those cells sort last in both directions. One such cell is deliberate: no source publishes a total system cost for the VR family, so that cell stays n/a rather than an estimate.
Bimanual workstations: fidelity first
The ALOHA design seats a human at a leader arm that is a twin of the follower arm the policy runs on. ACT, the policy built for that rig, learned six difficult bimanual tasks to 80-90% success, each from about 10 minutes of demonstrations, precisely because the demonstrations arrive already written in the robot's joint space Zhao 2023. Mobile ALOHA bolted a mobile base onto the same idea for whole-body tasks: with 50 demonstrations per task and co-training on existing static ALOHA data, success rates rose by up to 90% on tasks like sauteing shrimp and opening a two-door wall cabinet Fu 2024.
The hardware line behind these papers rebranded as Trossen AI in 2025-2026 and cut prices 30-34% across the board: $15,995 for the bimanual Stationary AI, $22,995 for Mobile AI on a base, $2,995 for the single-arm WidowX AI entry point, all running 500 Hz CAN FD control on the iNerve board with LeRobot and OpenPI integration Trossen Robotics 2026. The ALOHA 2 itself lists at $17,000-$32,000 depending on configuration in the LeRobot ecosystem pricing table alpibrusl 2026.
The trade is throughput. Each additional collector needs another full workstation and the space to put it, so fleet scale here is bought with capital, not with cheap duplication. No fleet-scale collection numbers are published for ALOHA-class rigs; the family earns its keep on data quality instead.
GELLO: the kinematic twin
GELLO takes the matched-kinematics idea and shrinks it. The device is a scaled replica of the target arm built from 3D-printed links and off-the-shelf motors, so the operator moves it like the arm itself and joint readings map one-to-one onto robot commands. Parts cost under $300, assembly takes about 30 minutes, and the published designs cover Franka, UR5, and xArm Wu 2023. In the paper's user study, 12 participants collected demonstrations for five bimanual UR5 tasks using GELLO, a VR controller, and a 3D spacemouse; GELLO came out more reliable and faster than both Wu 2023.
Two limits define the family. The $300 buys a controller only; the robot arm it drives is a separate cost. And kinematic matching is per-model: every new arm needs its own GELLO design, which is the price of the one-to-one joint mapping.
UMI: collection without a robot
UMI removes the robot from collection entirely. The operator holds a 3D-printed parallel-jaw gripper fitted with a wrist-mounted GoPro and a 155-degree fisheye lens, performs the task wherever it naturally happens, and SLAM recovers the gripper pose from the video Chi 2024. The paper bills the printed gripper at $73 and the GoPro plus accessories at $298, for a $371 collection rig Chi 2024. The project reports collection at about 30 seconds a demonstration: 111 demonstrations per hour on the cup arrangement task, measured against 35 per hour for spacemouse teleoperation and 231 per hour for the unassisted human hand Chi 2024.
The price of robot-free collection is an embodiment gap the interface has to close at deployment time. UMI does it with inference-time latency matching and a camera-relative action representation, and the resulting policies deploy zero-shot onto any arm with a parallel-jaw gripper of more than 85 mm stroke; the paper demonstrates UR5 and Franka Chi 2024. What the data lacks is a force channel, which matters for contact-rich tasks. Where robot-free collection fits in the wider data economics is the subject of The Data Bottleneck.
VR teleoperation: scale through headsets
VR teleoperation keeps the robot fixed and lets the operator move from inside a headset, with the robot mirroring the operator's hand or controller poses through a retargeting layer. DROID is the fleet-scale proof of the family: 50 operators across 13 institutions teleoperated Franka Panda rigs with Meta Quest 2 headsets and collected 76,000 trajectories totaling 350 hours over 12 months Khazatsky 2024. Diffusion policies trained on that pool beat policies trained on the next-best dataset by 22% in-distribution in controlled comparisons Khazatsky 2024, and the project maintains calibration protocols because multi-site retargeting drifts without them.
The newer systems push on immersion and fidelity. Open-TeleVision renders the robot's surroundings stereoscopically and mirrors the operator's arm and hand motion, validated on long-horizon tasks on two humanoid platforms Cheng 2024. Bunny-VisionPro runs on an Apple Vision Pro (launch price $3,499), adds low-cost haptic feedback devices for the operator, and builds collision and singularity avoidance into the retargeting Ding 2024. The family's weakness is the embodiment gap itself: human hands share no kinematics with robot arms, so every action passes through IK and calibration, and that mapping is where demonstration quality is kept or lost.
Choosing by axis
No rig wins all four axes, which is why all four families are still in use. If the constraint is data fidelity on one platform, an ALOHA-class workstation or a GELLO built for the target arm records demonstrations with the smallest gap between collection and deployment. If the constraint is demonstration volume across many scenes, UMI collects without a robot in the room and VR scales through commodity headsets, as DROID showed across 13 institutions. If the constraint is capital, GELLO at under $300 and UMI at $371 per gripper are the entry points, and the bimanual workstation tier starts at $15,995. The embodiment gap is the bill each family pays differently: matched kinematics up top, interface design (UMI) and calibration protocols (DROID) in the middle, retargeted human motion at the VR end.
Where this data goes next, and what it can and cannot teach a policy, continues in Major Datasets. The arms and sensors these rigs attach to are cataloged in Hardware Taxonomy, and what demonstrations become inside a policy starts at Behavior Cloning Foundations.