Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1)

TL;DR

mimic robotics announces a vertically-integrated dexterous-manipulation stack: the mimic hand M1 (15 active DoF, 21 total joints, tendon-driven, highly backdrivable with <0.05 Nm backdrive torque and 25+ kg payload) plus the mimic wearable U1 exoskeleton whose linkage mechanically enforces 1:1 kinematic + tactile + wrist-camera correspondence with M1, so operators wear the “robot’s skeleton” while collecting data. The core bet is a fixed-morphology data pyramid — human video at the base, wearable-collected demonstrations in the middle, on-robot teleop/deployment at the top — with the same hand kinematics used at every layer so there is no cross-embodiment gap between pretraining and deployment. Underneath sits mimic-ipc, a custom zero-copy real-time middleware (89 ns median latency on HD image payloads, ~18,000× speedup over FastDDS + shared memory) and a sampling-based retargeter released separately as smooth-operator. The post frames M1+U1 as the hardware substrate for their in-development Video Action Model line (following mimic-video).

Key claims

Two-finger grippers introduce a cross-embodiment gap that misaligns pre-training on human video with robot deployment: the end effector shape differs, and the discrete step-sequence a human uses is not the sequence a two-finger gripper needs [§Introduction]. mimic’s solution is to hold morphology constant across pretrain / demo-collection / deployment by using anthropomorphic hands at every phase.
The mimic hand M1 exposes 15 actuated + 6 coupled = 21 joints, ≥25 kg cylindrical-power-grasp payload, 25 N stretched-out fingertip force, <0.05 Nm joint backdrive torque, <0.1 N force-estimation sensitivity via dual motor/joint encoders, ±0.18 mm fingertip position accuracy, and <0.3° joint backlash [§mimic hand M1 – Specifications].
Tendon-driven wins over fully-integrated hands at industrial scale because forearm-housed motors deliver high continuous torque, backdrivability, and off-the-shelf drivetrain robustness that miniaturized in-hand motors cannot; fully-integrated hands are forced to trade backdrivability for payload via high gear ratios [§Challenging requirements to meet industrial needs].
Route every tendon over bearings/pulleys, not Bowden tubes: the capstan effect makes Bowden-driven friction and effective compliance vary with wrist configuration and drift with cycle count, producing dynamics that are hard to model; pulley routing gives linear transmission with predictable friction that stays consistent unit-to-unit — a precondition for policies to transfer between hands [§Challenging requirements to meet industrial needs].
mimic deliberately departs from strict anatomical fidelity at the wrist and forearm — human-matched geometry only matters near the fingertips where contact and demonstrations are generated, and the wider wrist/forearm is largely hidden behind the wrist-mounted camera so it doesn’t affect demonstration transfer [§Challenging requirements to meet industrial needs].
The mimic wearable U1 is a passive, user-actuated exoskeleton with 14 tracked + 6 coupled DoF, geometry-locked to M1 rather than adjustable, so demonstrators feel real contact forces through their own fingers, act without control latency or retargeting, and record demonstrations at near-natural human speed [§Scaling Data Beyond Teleoperation, §U1 Specifications].
Sensing is mirrored 1:1 between M1 and U1: tactile sensors, joint encoders, and wrist cameras are placed in one-to-one correspondence, so the demonstrator experiences the same modalities the robot will later see [§Scaling Data Beyond Teleoperation].
Teleoperation has structural scaling limits: latency, retargeting friction, operator discomfort at length, and the requirement for real factory hardware bound both throughput and the fidelity of any imitation-learned policy — motivating the wearable-first middle tier of the data pyramid [§Scaling Data Beyond Teleoperation].
mimic-ipc achieves ~89 ns median latency on HD image payloads across separate cores, a ~18,000× speedup over FastDDS-with-shared-memory, and 350× less jitter than ROS2 on a 232-byte control topic — the argument being that ROS2/LeRobot IPC bottleneck end-to-end imitation-learned policies more than model quality does [§mimic-ipc: Lightning fast and reliable, figure caption].
The teleop retargeter is sampling-based, not gradient-based, on the argument that gradient methods introduce noise artifacts on the 15-DoF M1 that classical model-based path planning avoids; a snapshot of the state of the art is released as smooth-operator [§A World-class Teleoperation Stack].
Observability and data collection are unified into one system built on mimic-ipc’s zero-copy primitives, with encoding/compression/persistence pushed to async processes — framed as a precondition for predictive maintenance and fleet lifecycle management [§Spotless and Zero-cost Telemetry].

Method

The stack has three co-designed layers.

Hardware — mimic hand M1: A tendon-driven anthropomorphic hand with 15 active DoF spread across 21 joints (including thumb opposition and finger abduction, matched to the human hand’s most functional DoF so demonstration data transfers cleanly). Motors sit in the forearm; tendons run over sized bearings and pulleys (not Bowden tubes) to give a linear, low-friction, high-stiffness transmission with rolling rather than sliding contact. Bi-directional actuation + full joint encoders + a dual motor/joint encoder scheme give unidirectional force resolution down to 0.1 N via motor current alone, augmented with dedicated fingertip tactile sensors that add tangential shear + higher spatial resolution at the contact points that matter most. The wrist and forearm are deliberately widened past anthropomorphic — the argument is that geometric fidelity only matters near the fingertips (where contact happens and where the wrist camera looks) and non-fidelity behind the wrist buys robust structural load paths + linear tendon routing.

Data-collection hardware — mimic wearable U1 (“umimic”): A passive exoskeleton whose rigid linkage mechanism mechanically constrains the user’s hand into the space of motions M1 can produce, enforcing 1:1 correspondence on every DoF. The thumb is the hard case (opposition + CMC kinematics) — solved by placing the user’s four fingers behind M1’s coupling mechanisms and their thumb above the robot thumb, trading a fixed geometry (only fits certain hand sizes) for an exact kinematic mapping. Tactile sensors, joint encoders, and wrist cameras are mounted at the same points as on M1, so demonstrator kinesthetic + visual + tactile observations match the robot’s data stream exactly. Contrast with UMI-style handheld end-effectors: mimic extends the UMI thesis (skip the robot at collection time) from two-finger grippers to a fully dexterous hand by fixing the wearable to the robot’s kinematics rather than the human’s.

Infrastructure — mimic-ipc + telemetry + retargeter: A custom hard-real-time zero-copy IPC middleware replaces ROS2 / LeRobot IPC in the entire stack. On HD image payloads it hits ~89 ns median latency across cores (compared to millisecond-scale jitter typical of FastDDS shared-memory setups) and 350× less jitter than ROS2 on a low-dimensional control topic. Telemetry reuses the same zero-copy primitives, running encoding and persistence in async processes so recording doesn’t steal cycles from control loops. Teleop retargeting uses a sampling-based solver (a snapshot of the state of the art is open-sourced at smooth-operator) rather than the gradient-based retargeters common in the literature, on the argument that gradient methods produce noise artifacts on high-DoF hands.

The M1+U1+infra combination is framed as the hardware substrate for mimic’s Video Action Model bet — an extension of their earlier mimic-video work — where the same hand morphology appears in the human video pretraining substrate (via the wearable) and at deployment.

Results

No learned-policy benchmark numbers are reported — this is a hardware + infrastructure system announcement, not a learning-paper. Reported measurements are hardware/infra characterizations:

mimic-ipc latency: ~89 ns median on HD image payloads, ~18,000× speedup over FastDDS with shared memory [§mimic-ipc figure caption]. On a 232-byte low-dimensional control topic, jitter is 350× lower than ROS2 FastDDS and stays in the nanosecond domain [§mimic-ipc].
M1 hardware: 15 active + 6 coupled DoF, ≥25 kg cylindrical-power-grasp payload, 25 N stretched-out fingertip force, <0.05 Nm backdrive torque, <0.1 N force-estimation sensitivity, ±0.18 mm fingertip position accuracy, <0.3° backlash, 15:1 minimum pulley:tendon diameter ratio, MCP up to 135°, hyperextension up to 45° [§Specifications].
U1 hardware: 14 tracked + 6 coupled DoF, ±0.18 mm fingertip position accuracy (via joint encoders), tactile sensing matching M1 [§Specifications].
smooth-operator retargeter: released as an open snapshot of mimic’s internal state of the art for high-DoF human-to-hand retargeting — no benchmark numbers in this post; benchmarking framework is deferred to the linked release.

Why it’s interesting

Fits directly into the Human-to-Robot Retargeting cluster’s most-active axis — where the embodiment-invariant interface lives. Existing filed positions place that interface at a joint-pose retargeter with physics-in-the-loop (ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting), a 4D hand-object reconstruction (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos), a reduced 3-DoF wrist-translation bridging signal (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), or a task-intent abstraction (LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition). mimic stakes out a hardware-first position: fix the embodiment-invariant interface at the level of the hand kinematics itself by shipping a wearable that mechanically constrains the human into the robot’s motion space, eliminating the retargeting problem at collection time. This is the same shortcut Xiaomi took at the end-effector level with hand-held UMI grippers in Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training, extended one embodiment upward from two-finger to dexterous.

Complements UMI-Bench 1.0: An Open and Reproducible Real-World Benchmark for Tabletop Robotic Manipulation with UMI Data (which benchmarks policies trained on the original two-finger UMI data) by generalizing the “collection tool is the retargeting target” thesis to a full anthropomorphic hand. The Tactile sensing for manipulation cluster’s Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks finding that placement dominates sensor type (whole-hand coverage > fingertip-only) makes M1’s fingertip-only tactile augmentation (on top of joint-encoder force sensing) a live design question: mimic’s argument is that motor-current + dual-encoder gives implicit force sensing on every joint, so the effective coverage is whole-hand — but no head-to-head with dedicated taxels on the palm and proximal phalanges is offered.

Structurally distinct from the VLA Models cluster’s dominant recipe (pretrain a VLM on internet data, add an action head, teleoperate the last mile). mimic’s argument is that teleop is structurally-capped by human latency + retargeting friction + operator discomfort, and that only wearable-collected data at wearable-quality can escape those caps — an infrastructure position rather than an algorithm position.