Robotics

Mimic Hand M1: Swiss Tendon Hand Built for Factory AI

Shar Hendrix 4 min read

Swiss startup Mimic Robotics unveiled the mimic hand M1, a tendon-driven industrial hand with actuators in the forearm and a matching wearable for collecting human demonstrations. Company blog and Interesting Engineering coverage put the design squarely on factory dexterity, not lab theater.

The M1 is five-fingered, made in Switzerland, and sold as part of a full stack: hand hardware, the U1 (“umimic”) exoskeleton, and in-house real-time middleware.

Pair of Mimic robotic hands in mesh gloves manipulating a small object
Mimic hands in mesh work gloves handling a small object. Source: Interesting Engineering / Mimic YouTube.

Specs that matter

From Mimic’s product announcement:

SpecM1 claim
Active DoF15 actuated + 6 coupled = 21 total
Payload> 25 kg steady-state cylindrical power grasp
Fingertip force25 N (stretched out)
Backdrive< 0.05 Nm (senses ~50 g via motor current)
Force sensitivity< 0.1 N (dual encoders)
Position accuracy± 0.18 mm (closed loop, joint encoders)
Tactile tipsNormal force, tangential shear, multi-point contact
WeightAbout 1.8 kg (4 lb) per Interesting Engineering

Tendons run over bearings and pulleys, not Bowden tubes, to keep friction low and predictable across wrist angles. Mimic widened the wrist/forearm relative to human anatomy so the routing stays linear behind a wrist camera.

Interesting Engineering also notes a YouTube launch clip (ikjPRgE8WLM).

Mimic Hand M1 launch footage. Source: Mimic Robotics / YouTube.

Why the wearable exists

The U1 passively couples a human hand to M1-matched kinematics (14 tracked + 6 coupled DoF). Sensors and a wrist camera sit in the same places as on the robot, so demos share morphology with deployment. Mimic’s argument: pre-train on human video, mid-tier wearable data, top-tier robot teleop—without jumping from human fingers to a two-finger gripper mid-pipeline.

They also ship mimic-ipc, a zero-copy real-time bus they claim hits ~89 ns median latency on HD image payloads between processes—orders of magnitude under ROS2 FastDDS in their own charts.

Mimic data pyramid for human video, wearable, and robot teleop
Mimic’s data pyramid: human video at the base, wearable in the middle, robot teleop at the top. Source: Mimic Robotics blog.

A Human’s Take

I’m so here for a hand that treats force sensing as a first-class design requirement instead of a fingertip sticker after the fact. Sub-0.05 Nm backdrive plus a matched wearable is the kind of boring infrastructure that makes imitation learning less of a retargeting nightmare. Next receipt I want: cycle-time and failure rates on a real shift, not just another desk-demo power grasp.

Sources