Dexterous humanoid hands

Counting what a hand can actually do

AGIBOT's OmniHand SDK supports a gripper with 1 degree of freedom (OmniPicker 3), a lightweight hand with 4 degrees of freedom (OmniHand 3 Lite), a professional hand with 12 degrees of freedom (OmniHand Pro 2025) and the OmniHand 2025, which has 10 active and 6 passive degrees of freedom. ^[1]

Fourier offers two hand options on one platform. The support page for the Fourier GR-3 lists a base hand with 6 degrees of freedom and an optional 12-DOF dexterous hand. ^[10] Fourier's ActionNet dataset likewise records two types of Fourier dexterous hands, with six and twelve degrees of freedom. ^[4]

The HumDex paper describes a setup with two 20 DoF dexterous hands, with 4 independent degrees of freedom per finger. ^[6]

When you compare hands, compare actuated joints, and read a listing carefully for passive joints: the AGIBOT SDK lists the OmniHand 2025 with 10 active and 6 passive degrees of freedom. ^[1]

Why some makers do not build a five-finger hand

Agility Robotics says that instead of developing a delicate five-fingered hand that mimics human morphology, it built a modular system, which means swappable grippers. ^[7] Galbot lists the end effector of the Galbot G1 as swappable between dexterous hands, suction cups and grippers. ^[11]

Humanoid describes the end-effectors of the HMND 01 Alpha Bipedal as a modular design featuring a 12-DOF 5-finger hand or a 1-DOF parallel gripper. ^[12]

Boston Dynamics names gripper design for Atlas as an iteration target, aiming for the right mix of strength, fine motor skill and durability. ^[8]

Tactile and close-range sensing

In its OmniHand SDK documentation, AGIBOT describes the OmniHand 2025 as weighing only 500g, with 400+ tactile points and 0.1N array resolution, and a maximum fingertip force of 5N. ^[1] The documentation also says that the OmniHand 3 Lite exposes raw tactile streaming through its ROS2 node, and that the OmniPicker 3 gripper has two-region tactile sensing. ^[1]

Figure says each hand of Figure 03 has an embedded palm camera with a wide field of view and low-latency sensing, which gives redundant, close-range visual feedback during grasps. ^[9] Figure also says its fingertip sensors detect forces as small as three grams of pressure, and that softer fingertips increase the contact area and so give more stable grasps across objects of varied shapes and sizes. ^[9]

The RoboTacDex dataset, built on the Unitree G1, provides multi-view RGB and depth information together with tactile feedback and semantic annotations. ^[5] If you plan to train on touch, check that the hand exposes it through a documented interface and that your recording format keeps it.

Retargeting human hand motion to a robot hand

The dex-retargeting library provides various retargeting optimizers that translate human hand motion to robot hand motion, and its repository is derived from the AnyTeleop project. ^[2] One practical warning from its documentation is that URDF parsers, simulators and robot drivers can parse a URDF with different joint orders, so you should handle joint order explicitly by joint name. ^[2]

The HumDex paper describes the common optimisation-based approach as a constrained inverse kinematics or nonlinear least-squares problem. ^[6] HumDex instead trains a lightweight MLP regressor that maps the 3D positions of the operator's five fingertips, captured with inertial gloves, to the 20-DoF joint angles of the robot hand. ^[6] Its training pairs come from an offline optimisation-based retargeting process, and the calibration set is about 20k frames, under 20 minutes of recording. ^[6]

In the paper's closed-loop trials with gloves, the learned retargeter succeeded in 20 of 30 trials on Scanner Triggering against 14 of 30 for optimisation, and in 29 of 30 against 23 of 30 on Doll Grasping. On Hanger Stabilization the figures were 20 of 30 against 21 of 30. ^[6]

Data, teleoperation and policy learning

The RoboTacDex dataset consists of 6k trajectories covering 19 tasks, 23 skills and interactions with 22 objects. ^[5]

The HumDex paper describes a teleoperation setup built on inertial motion capture, with inertial gloves for hand tracking. ^[6] It says that vision-based and VR-based systems suffer from occlusion because operators must keep their hands visible. ^[6] The paper compares HumDex with a baseline teleoperation system. It reports that the baseline was unable to perform the Scan&Pack task because of severe occlusion. On the four tasks that both systems completed, collecting 60 episodes took 44.3 minutes with HumDex against 59.8 minutes for the baseline. Teleoperation success was 91.7% against 74.6%, and trained policies reached 80.0% against 57.5%. ^[6]

HumDex also trains in two stages: first on human data, then on robot data. A naive mix of human and robot data failed to converge and reached 0% success. ^[6] The paper reports 21 of 30 successes on unseen object positions for the two-stage policy against 12 of 30 for a robot-data-only policy. ^[6] The HumDex code repository documents the pipeline, including ACT training and the teleoperation scripts. ^[3]

Questions to ask when evaluating a hand

A short checklist follows from the sources above.

• How many joints are actuated, and how many are passive? The OmniHand 2025 listing separates 10 active from 6 passive degrees of freedom. ^[1]
• Does the hand report touch or only position? Check what the SDK exposes, such as the raw tactile streaming of the OmniHand 3 Lite through its ROS2 node. ^[1]
• Is a simpler end effector enough? Galbot lists the end effector of the Galbot G1 as swappable between dexterous hands, suction cups and grippers. ^[11]
• How will human motion reach the hand? Check whether a retargeting method exists for your hand, and whether it handles joint order by name. ^[2]
• How strict is the task criterion? HumDex required five consecutive scanner trigger actuations, each held for 3 seconds, rather than a single tap. ^[6]