Agibot OmniHand is designed for gesture interaction, touch interaction, and light-duty manipulation. It provides 10 active degree sof freedom across 16 total degrees of freedom, with three active DoFs in the thumb, two in the index finger, one in the middle finger, two in the ring finger, and two in the little finger. The tactile version also provides force sensing across the fingertips, palm, and back of the hand.
MANUS gloves serve as the human motion interface for OmniHand teleoperation. MANUS gloves track the fingertip position and rotation, and MANUS Core further process to 25 degrees of freedom anatomical hand skeleton, along with finger joint angles. A retargeting layer then converts this human hand data into commands for OmniHand's 10 active motor axes.
OmniHand does not independently control every physical finger joint. Several joints are mechanically coupled, and the distribution of active degrees of freedom differs from the anatomical hand model captured by MANUS. A direct joint-to-joint mapping would therefore not represent the same motion across the two embodiments.
Instead, the retargeting system can use MANUS fingertip positions and rotations as its primary control targets. This allows the system to focus on where each fingertip should move, rather than attempting to reproduce every human joint angle on the robot.
Inverse kinematics or constrained optimization can then determine an achievable OmniHand configuration within its active control space and joint limits. The objective can prioritize fingertip placement, thumb opposition, grasp geometry, and the relative distance between the fingers.
MANUS data also provides the wrist pose and complete hand skeleton surrounding these fingertip endpoints. This gives researchers additional context for interpreting the intended movement and developing custom retargeting algorithms, cross-embodiment policies, or data collection pipelines.
The tactile OmniHand configuration provides one-dimensional force sensing across the fingertips, palm, and back of the hand, with a specified resolution of 0.1 N and sensing range of 0 to 20 N.
Within a custom bilateral teleoperation loop, these contactsignals can be mapped to MANUS Metagloves Pro Haptic. Feedback on thecorresponding human finger can help the operator recognize contact and adjustfingertip placement, hand pose, or grip intensity during manipulation.
A MANUS and OmniHand integration can record human fingertip trajectories, wrist motion, retargeted robot commands, OmniHand state feedback, tactile measurements, camera observations, and operator corrections within the same demonstration.
This synchronized data can support imitation learning, tactile representation learning, contact prediction, foundation models, and world models for dexterous manipulation. MANUS provides the structured human motion data required to connect fingertip-based OmniHand teleoperation with broader robot learning pipelines.
