Mimic Robotics is a Swiss robotics company that develops dexterous robot hands, human demonstration equipment, control software, and learned manipulation policies for industrial work. The company was founded in 2024 as a spin-off from ETH Zurich's Soft Robotics Lab.[3][4] Its stated technical focus is general-purpose hand manipulation rather than a complete humanoid robot.

In July 2026, Mimic Robotics announced the mimic hand M1 and the mimic wearable U1, also called umimic. The hardware is part of a vertically integrated physical AI system that also includes data collection, teleoperation, middleware, model training, and on-device inference.[2] One week later, the company and Black Forest Labs presented FLUX-mimic, a preview Video-Action Model built on the multimodal FLUX 3 backbone.[1][10]

Company development

Mimic Robotics originated in research at ETH Zurich's Soft Robotics Lab. Venture Kick identified Stefan Weirich, Elvis Nava, Stephan-Daniel Gravert, Benedek Forrai, and professor Robert Katzschmann as the early team behind the spin-off.[4] The ETH AI Center later described Weirich as chief executive, Nava as chief technology officer, Gravert as chief product officer, and Katzschmann as a co-founder and scientific adviser.[6]

The company received CHF 150,000 from Venture Kick in February 2025.[4] In November 2025, it raised a US$16 million seed round led by Elaia, with Speedinvest and other investors participating. Venture Kick reported that the round brought total funding to more than US$20 million.[5] Mimic Robotics said the funding would support its models, robotic hands, and industrial deployments, while the ETH AI Center reported pilot work in manufacturing and logistics.[5][6]

The company's research and product path can be divided into four public stages:

Date	Development	Scope
2024	Company formation	ETH Zurich recognized Mimic Robotics as a spin-off from its Department of Mechanical and Process Engineering.[3]
June 2025	mimic-one	A research system pairing a 16-DoF tendon-driven hand with a Franka Emika Panda arm and a diffusion-based control policy.[7]
December 2025	mimic-video	A Video-Action Model that used latent features from a pretrained video generator to condition a robot action decoder.[8][9]
July 2026	M1, U1, and FLUX-mimic	Announced hardware and software stack, followed by a preview model developed with Black Forest Labs.[1][2]

From mimic-one to Video-Action Models

The 2025 mimic-one preprint described an earlier 16-degree-of-freedom research hand with 20 joints, antagonistic tendon actuation, silicone contact surfaces, and two wrist cameras. Researchers mounted it on a seven-DoF Franka Emika Panda arm and collected demonstrations through motion-capture gloves or Apple Vision Pro tracking. The paper evaluated bread handling, bottle sorting, and battery insertion, and reported that data diversity, filtering, and explicit recovery demonstrations improved task success.[7]

That hand should not be conflated with the later M1. The research prototype had 16 active DoF and a reported power-grasp payload above 7 kg. The July 2026 M1 specification instead lists 15 actuated and six coupled DoF, with a claimed steady-state power-grasp payload above 25 kg.[2][7]

Mimic-video changed the model design rather than the hand alone. Conventional vision-language-action models commonly start with a vision-language backbone and train an action component on robot trajectories. Mimic-video instead used a pretrained video generator to form latent predictions of future visual states. A smaller inverse-dynamics decoder translated those latent features into low-level actions.[8] The published implementation used a 2-billion-parameter NVIDIA Cosmos-Predict2 backbone and released code and checkpoints for the Bridge and LIBERO settings.[9]

The mimic-video authors reported up to 10 times greater sample efficiency and twice the convergence speed of their VLA comparison. They also evaluated a bimanual setup on package sorting and tape stowing.[8][9] These figures came from the authors' experiments on their selected simulated and real tasks. They did not constitute an evaluation of the later FLUX-mimic model.

mimic hand M1

The mimic hand M1 is a five-fingered robotic hand announced for industrial manipulation. Its motors sit in the forearm and drive the finger joints through bidirectional tendons routed over pulleys and bearings. Mimic Robotics chose this layout to allow larger actuators than would fit inside the palm, while retaining backdrivability and reducing sliding friction in the tendon path.[2] Independent coverage of the launch showed company demonstrations involving tweezers, electronic components, bolts, and hand signals.[13]

The hand includes encoders at the motors and joints, plus fingertip tactile sensing for normal force, tangential force, and contact location. Mimic Robotics says motor-current measurements can also estimate contact force. The following values are manufacturer specifications, not results from an independent laboratory test.[2][14]

M1 characteristic	Published specification
Actuation	Bidirectional, pulley-guided tendon drive
Degrees of freedom	15 actuated + 6 coupled = 21 total
Steady-state payload	More than 25 kg in a cylindrical power grasp
Fingertip steady-state force	25 N with fingers stretched
Joint backdrive torque	Less than 0.05 Nm
Force-estimation sensitivity	Less than 0.1 N
Closed-loop fingertip position accuracy	+/-0.18 mm
Joint backlash	Less than 0.3 degrees
Fingertip sensing	Normal force, tangential force, and multi-point contact location
Cameras	Synchronized global-shutter cameras

The company departed from strict anatomical proportions around the wrist and forearm to make room for the linear tendon routing and industrial components. It retained a human-like arrangement at the fingers and thumb, where contact and demonstration data are produced.[2] Price, order terms, and general commercial availability were not disclosed with the announcement.[14]

mimic wearable U1

The mimic wearable U1, nicknamed umimic, is a passive exoskeleton for collecting human manipulation demonstrations without operating a robot during every recording. A rigid linkage constrains the wearer's fingers to motions that the M1 can reproduce. The device also places joint encoders, tactile sensors, and a wrist camera in positions intended to correspond to the robot's observations.[2]

Mimic Robotics describes the mapping as 1:1 at the level of hand kinematics and contact geometry. The published tables nevertheless list 14 tracked plus six coupled DoF for U1, or 20 total, compared with 21 total DoF for M1. The company does not explain the one-DoF difference in the announcement. U1's geometry is fixed rather than adjustable to every user, so it only fits a limited range of hand sizes. Mimic Robotics presents that constraint as a tradeoff for direct mechanical coupling, contact feedback through the wearer's fingers, and the absence of software retargeting latency.[2]

The U1 sits in the middle of the company's three-level data strategy. Ordinary human video forms the broad base, U1 demonstrations add robot-compatible motion and sensor readings, and teleoperation or autonomous deployment produces the smaller amount of data collected directly on robots. The goal is to keep hand morphology similar across pretraining, demonstration collection, and robot execution, instead of transferring human-hand observations to a two-finger gripper.[2]

Control and data infrastructure

M1 and U1 are accompanied by software for control, recording, teleoperation, and inference. Mimic Robotics calls its inter-process communication layer mimic-ipc. It uses a zero-copy design to move camera, sensor, and action data between processes with low latency and jitter. The same primitives support asynchronous telemetry and dataset recording.[2]

The stack also includes a hand-pose retargeter for conventional teleoperation. U1 is intended to avoid retargeting during wearable data collection, but direct robot teleoperation remains part of the system for deployment support, failure handling, and task-specific data. Mimic Robotics has published a snapshot of its retargeting benchmark, while its broader middleware remains a company-developed component.[2]

FLUX-mimic

FLUX-mimic is a preview robot-control model announced on July 23, 2026. It uses the video path of FLUX 3, a multimodal foundation model trained across images, video, and audio. FLUX 3 was itself in early access at the time.[10][11] Mimic Robotics trained the backbone further on robot-manipulation data and attached a compact action decoder to its internal representation of a predicted future.

The design treats the video backbone as a learned world model, but the deployed policy does not render a full video before each movement. The action decoder attends to intermediate latent features and produces chunks of robot actions after one backbone forward pass. Mimic Robotics also described quantization, cached context, and overlapping action prediction and execution as parts of its on-device implementation. The company reported running the model locally on a single NVIDIA RTX 5090.[1]

FLUX-mimic extends the earlier mimic-video recipe but uses FLUX 3 instead of Cosmos-Predict2. Black Forest Labs said it gave Mimic Robotics early access to FLUX 3 and worked with the company on the action modality and deployment system.[10] Mimic Robotics combines ordinary video, U1 recordings, and data from robot teleoperation or operation when training the system. It uses imitation learning for demonstrations and applies reinforcement learning afterward to refine policies on target tasks.[1]

Evaluation and deployment status

Mimic Robotics reported a 95% completion rate for a preview version of FLUX-mimic on an internal soft-body kitting task. In the same comparison, an adapted pi0.5 model trained on the company's data mix achieved 55%, and a flow-matching policy heavily post-trained on that task achieved 70%.[1] Black Forest Labs separately described 20-trial internal comparisons and reported that fine-tuning the video backbone with the action decoder improved results.[10] No independent reproduction, public FLUX-mimic checkpoint, or reproducible benchmark package accompanied the preview.

Mimic Robotics and Black Forest Labs said they had tested the system with Audi on production-related assembly and soft-material handling tasks.[1][10] Reporting by the Association for Advancing Automation confirmed the partnership through an interview with Mimic Robotics CTO Elvis Nava, but the performance and task-learning figures in that report remained company claims.[12] The cited work establishes a factory partner evaluation; it does not by itself establish fleet-scale commercial deployment.

FLUX-mimic was presented as a preview rather than an open release. Mimic Robotics discussed future interfaces based on goal images, motion sketches, and video demonstrations, but stated that these were planned for later versions.[1] As announced, the substantiated system consists of an experimental Video-Action Model, the M1 and U1 hardware, and the company's data and control infrastructure.

References

Mimic Robotics. "Introducing FLUX-mimic: Scaling Video-Action Models for General Purpose Dexterity." July 23, 2026. https://www.mimicrobotics.com/blog/introducing-flux-mimic
Mimic Robotics. "Solving Dexterity: A Full-Stack Approach." July 16, 2026. https://www.mimicrobotics.com/blog/solving-dexterity-a-full-stack-approach
ETH Zurich Department of Mechanical and Process Engineering. "Spin-offs." Accessed July 2026. https://mavt.ethz.ch/the-department/forschungstransfer/spin-offs.html
Venture Kick. "Mimic Robotics receives CHF 150,000 from Venture Kick to build embodied intelligence." February 20, 2025. https://www.venturekick.ch/Mimic-Robotics-receives-CHF-150000-from-Venture-Kick-to-build-embodied-intelligence
Venture Kick. "USD 16 million for Mimic Robotics to bring physical AI to industry." November 3, 2025. https://www.venturekick.ch/USD-16-million-for-Mimic-Robotics-to-bring-physical-AI-to-industry
ETH AI Center. "mimic: Developing Human-Level Dexterity for the Next Era of Industrial Automation." November 12, 2025. https://ai.ethz.ch/news-and-events/ai-center-news/2025/09/mimic_Robotics_Seed_Round_2025.html
Nava, Elvis, et al. "mimic-one: a Scalable Model Recipe for General Purpose Robot Dexterity." arXiv:2506.11916. June 2025. https://arxiv.org/abs/2506.11916
Pai, Jonas, et al. "mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs." arXiv:2512.15692v2. December 2025. https://arxiv.org/abs/2512.15692
mimic-video authors. "mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs." Project page. https://mimic-video.github.io/
Black Forest Labs. "FLUX 3 x mimic: The Next Generation of Video-Action Models." July 23, 2026. https://bfl.ai/blog/flux-3-mimic
Black Forest Labs. "FLUX 3 - Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence." July 23, 2026. https://bfl.ai/blog/flux-3
Szkutak, Rebecca. "Mimic and Black Forest Team to Train Robots." Association for Advancing Automation, July 23, 2026. https://www.automate.org/ai/industry-insights/mimic-and-black-forest-team-to-train-robots
Dotson, Kyt. "Mimic Robotics launches highly capable robotic hand that emulates human movements." SiliconANGLE, July 16, 2026. https://siliconangle.com/2026/07/16/mimic-robotics-launches-highly-capable-robotic-hand-emulates-human-movements/
RoboZaps. "mimic hand M1." Checked July 17, 2026. https://robozaps.com/products/mimic-hand-m1

Editing: Mimic Robotics