Skip to content
English
  • There are no suggestions because the search field is empty.

Mapping a Human Hand to a Robotiq 2F-85 Gripper for AI Learning

Attach a virtual 2F-85 to a detected hand armature to turn videos of human hands into gripper poses and openings.

01_hero

Physical AI needs demonstrations of manipulation, and people already record thousands of hours of hands doing useful work. The challenge is that a robot does not have a hand: a Robotiq 2F-85 has two parallel fingers, one opening value and a pose. For an AI, working out which gripper configuration matches a given human hand configuration is not obvious.

Attaching a virtual 2F-85 to the hand does that translation for you. For every frame of a video, you get:

  • the gripper pose (position and orientation)
  • the gripper opening, from 0 to 85 mm

This turns videos of people at work into data an AI can learn from, which makes it easier to teach a robot a task by showing it how a person does it.

What is a Hand Armature?

A hand armature represents the human hand as a simple skeleton: its bones become segments, and its articulations become joints between those segments. Each joint is labelled with a letter in the figures of this article.

02_hand_armature

Detecting the Hand Armature

Several types of tools can detect the configuration of the hand armature in space, relative to a reference frame. Their precision depends on the detection method, and choosing the most precise one is outside the scope of this article. Here is a non-exhaustive overview:

Source What it gives Notes
Ordinary video (webcam, phone, existing footage) Hand joints in the image; some tools also estimate a full 3D hand Works on videos that already exist; some tools run in real time on a CPU
VR / XR headset with hand tracking 3D hand skeleton from the headset cameras Headsets that follow the standard XR hand model give 26 joints per hand
Hand-tracking camera 3D hand skeleton Stereo infrared, on a desk or mounted on a headset
Depth camera Measured depth, combined with a 2D hand detector Depth is measured, not estimated
Data glove Finger joint angles measured on the hand No camera needed; keeps working when fingers are hidden
Optical motion capture 3D marker positions, fitted to a hand Several cameras and reflective markers; the most accurate option, used as a lab reference
Attaching the 2F-85 to the Hand Armature

For convenience, some geometry is added on top of the armature:

  • A is the middle of D and E.
  • G is the middle of B (index tip) and C (thumb tip).
  • The triangle EDG gives the plane used to orient the gripper.

03_added_geometry

The 2F-85 frame is placed at the gripper base: z runs along the gripper axis toward the fingertips, x is the closing direction, and y is normal to the plane of the fingers. The TCP sits between the pad tips.

04_gripper_frame

Once the armature is detected, place the virtual gripper with these four rules:

  1. Make the gripper XZ plane parallel to the EDG plane.
  2. Match the tip of the thumb-side gripper pad with the tip of the thumb (C).
  3. Keep an angle of 110° between the gripper axis (z) and the D to E direction.
  4. Place B in the plane of the index-side gripper pad. This sets the gripper opening.

05_gripper_attached

Tip: the 110° angle may need adjusting. It worked well with the hand armature from a VR headset's hand tracking, so it is a good value to start with.

If the hand opening (CB) is larger than the maximum gripper opening, the index tip ends up outside the plane of the gripper pad.06_index_tip_off_pad

Limitations of the Method

The method makes deliberate trade-offs to keep the gripper orientation stable, because a robot cannot follow sudden orientation changes:

  • The index tip does not match the tip of the gripper finger. Forcing them to match would make the gripper orientation unstable, especially when the fingers are close to each other. In practice this matters little: when you close your fingers, the index and thumb tips naturally meet, so a hand configuration where they do not is rare.

07_index_tip_not_matching

  • The gripper can appear to "fly above the hand". When the thumb tip is not in the DEG plane, the gripper looks attached to the thumb only. Placing the gripper in the CBA plane would follow the fingers more closely, but it makes the gripper unstable around its Z axis when the fingers are close together. This only happens when the thumb is in an uncomfortable position, which is rare.

08_flying_above_hand

  • The method assumes a grasp between the thumb and the index. This covers most of the way people pick objects, but humans also use more sophisticated grasps. For those, the method still produces a gripper configuration, but it may not be suited to performing the same manipulation.
Examples

09_example_webcam

Gripper attached to hands detected on a webcam video.

10_example_vr_headset

Gripper attached to the hand armature from a VR headset.

Conclusion

Attaching a virtual 2F-85 to a detected hand armature turns any recording of human hands into gripper poses and openings, a simple bridge between how people manipulate objects and how a robot can reproduce it. To see how Robotiq components fit a physical AI application, read Order preparation application with Robotiq components, or check out the 2F-85 adaptive gripper.

Need help? Contact Robotiq Support.

Updated: November 18, 2026