- Direct drive at every joint Motors drive the joints without tendons or high-ratio gearboxes, and each actuator's encoder reports the real joint state.
- 21 active degrees of freedom Every one of its 21 degrees of freedom is driven, in a 630 g hand measuring 205 × 93 × 50 mm.
- Vision-based tactile fingertips Fingertip sensors read 3D force vectors from 0 to 30 N at 0.005 N resolution and 0.04 mm spatial resolution, and they recognise texture.
- 300-point tactile palm A 3D tactile sensor with 300 tactile pixels covers the palm at 0.1 N resolution.
- ±0.1 mm, 0.3 s, 5 kg Repeatability holds to ±0.1 mm across the workspace, the hand opens or closes in 0.3 s and it carries 5 kg.
- Built for continuous duty AgiBot rates it at a 40–45 °C surface temperature under continuous load and validates it for 500,000 cycles.
- Isaac, MuJoCo and ROS 2 C++ and Python SDKs, ROS 2 reference nodes and native NVIDIA Isaac and MuJoCo assets come with it, over CAN FD or Ethernet.
Description
The AgiBot OmniHand 3 Ultra-M is AgiBot's flagship direct-drive dexterous robotic hand: 21 active degrees of freedom in a human-scale hand of 630 g, measuring 205 × 93 × 50 mm. Every joint is driven directly, without tendons or high-ratio gearboxes, and its fingertips register contact through vision-based tactile sensors.
AgiBot designed it for embodied-AI training, teleoperation and demonstration capture, and for bimanual, contact-rich manipulation research. Because each actuator's encoder reports the real joint state, the data a controller or learned policy works from matches what the hand actually did, with ±0.1 mm repeatability across the full workspace. AgiBot rates it for continuous duty, with a 40–45 °C surface temperature under continuous load and 500,000 validated cycles.
Each fingertip carries a vision-based tactile sensor that measures three-dimensional force vectors and the 6D resultant force from 0 to 30 N, at 0.005 N force resolution, 0.04 mm spatial resolution and 0.08 mm deformation accuracy, enough to tell slip from contact and an edge from a surface, and it recognises texture. The palm adds a 3D tactile sensor with 300 tactile pixels at 0.1 N. The hand carries a 5 kg payload and opens or closes in 0.3 s.
It connects over CAN FD or Ethernet. AgiBot provides C++ and Python SDKs with full API coverage, ROS 2 reference nodes for teleoperation and joint-space control, and native NVIDIA Isaac and MuJoCo assets for simulation and reinforcement-learning training, and builds the hand as a left and a right hand for two-handed robots. Below it in the family sit the OmniHand Pro 2025, with 12 active degrees of freedom and three-axis fingertip force, and the compact OmniHand 3 Lite; compare all our dexterous robotic hands.
Key Capabilities
- Direct drive at every joint: Motors drive the joints without tendons or high-ratio gearboxes, and each actuator's encoder reports the real joint state.
- 21 active degrees of freedom: Every one of its 21 degrees of freedom is driven, in a 630 g hand measuring 205 × 93 × 50 mm.
- Vision-based tactile fingertips: Fingertip sensors read 3D force vectors from 0 to 30 N at 0.005 N resolution and 0.04 mm spatial resolution, and they recognise texture.
- 300-point tactile palm: A 3D tactile sensor with 300 tactile pixels covers the palm at 0.1 N resolution.
- ±0.1 mm, 0.3 s, 5 kg: Repeatability holds to ±0.1 mm across the workspace, the hand opens or closes in 0.3 s and it carries 5 kg.
- Built for continuous duty: AgiBot rates it at a 40–45 °C surface temperature under continuous load and validates it for 500,000 cycles.
- Isaac, MuJoCo and ROS 2: C++ and Python SDKs, ROS 2 reference nodes and native NVIDIA Isaac and MuJoCo assets come with it, over CAN FD or Ethernet.
This model is offered on a quotation basis. Configuration, accessories, integration scope and delivery schedule all change the final figure, so pricing is confirmed per project rather than listed.
Buying through Robotix Market
This unit is sourced new through established distribution channels and ships worldwide with the manufacturer warranty. Technical support is coordinated with the manufacturer's own service network. Deployment scope, configuration, and commissioning responsibilities are agreed as part of the quotation — on-site installation is not included in the product price.
Features
Specs
| Brand | AgiBot |
|---|---|
| Model | OmniHand 3 Ultra-M (H3UM) |
| Hand side | Left or right; state the side, or a pair, with your order |
| Fingers | 5 |
| Degrees of freedom | 21 active |
| Drive | Direct drive at every joint; joint state read from each actuator's encoder |
| Weight | 630 g |
| Size | 205 × 93 × 50 mm |
| Payload | 5 kg |
| Opening / closing time | 0.3 s (typical) |
| Repetitive positioning accuracy | ±0.1 mm across the full workspace |
| Fingertip tactile sensing | Vision-based tactile sensors: texture recognition, 3D distributed force and 6D resultant force |
| Fingertip force range | 0–30 N |
| Fingertip force resolution | 0.005 N |
| Fingertip spatial resolution | 0.04 mm |
| Deformation field accuracy | 0.08 mm |
| Palm tactile sensing | 3D tactile sensor, 300 tactile pixels, 0.1 N force resolution |
| Continuous duty | 40–45 °C surface temperature under continuous load, without throttling |
| Validated cycles | 500,000 |
| Communication interfaces | CAN FD, Ethernet |
| SDK | C++ and Python, full API coverage |
| ROS 2 | Reference nodes for teleoperation and joint-space control |
| Simulation | Native NVIDIA Isaac and MuJoCo assets; supports reinforcement-learning training |
| Operating voltage | Confirm at quote |
| Robot mounting | Confirm at quote |
Industries
Frequently Asked Questions
It is offered by quotation — configuration, quantity and delivery are confirmed per project.
Twenty-one, all of them actively driven, each joint by its own direct-drive actuator.
The motors drive the joints directly instead of through tendons or high-ratio gearboxes, and each actuator's encoder reports the joint state. Controllers and learned policies therefore work from the hand's real position, with ±0.1 mm repeatability.
Vision-based tactile sensors in the fingertips measure 3D force from 0 to 30 N at 0.005 N resolution and recognise texture, and a 3D tactile palm has 300 tactile pixels at 0.1 N resolution.
C++ and Python SDKs, ROS 2 reference nodes for teleoperation and joint-space control, and native NVIDIA Isaac and MuJoCo assets. It connects over CAN FD or Ethernet; left and right hands are available, and fitment to your robot is confirmed with the quotation.
The Ultra-M has 21 active degrees of freedom against 12, direct drive instead of a linkage and screw rod, and vision-based tactile fingertips at 0.005 N against three-axis force sensing at 0.01 N, at 630 g against up to 750 g.
Recently Viewed
You haven't viewed any products yet.
