Solving Dexterity: A Full-Stack Approach

Today we are proud to announce the mimic hand M1, our highly backdrivable, tendon-driven robotic hand designed and built in-house for industrial automation. Alongside it, we are introducing the mimic wearable U1 (“umimic”), our exoskeleton that captures human demonstrations directly matched to the kinematics of mimic hand M1. Together with our fully in-house developed software stack, these systems provide an integrated platform to enable general-purpose dexterous manipulation.

From day one, mimic has been laser-focused on the single goal of general-purpose dexterous manipulation. We believe that the only way to solve general-purpose manipulation while creating value at scale is to take the “full-stack approach”. Deploying autonomous fleets of robots will depend on vertically integrated stacks with full hardware observability, just as full vertical integration mattered in autonomous driving.

Every serious attempt to bring foundation models to the physical world runs into the same wall: robotics has no equivalent to internet-scale training data. So an optimal approach must draw on multiple sources and combine data across qualities and quantities. What pre-training delivers is representations - visual, semantic, behavioral, and physical. The largest source of behavioral and physical priors at meaningful scale is large-scale video of humans. For post-training, nothing beats the robot: live deployment data and teleoperation, collected on the target hardware in the target environment.

However, the embodiments of traditional two-finger gripper robots and the human embodiment are fundamentally mismatched. If we ground pre-training in human video and then deploy a two-finger gripper, we introduce a cross-embodiment gap that renders our pre-training and post-training phases misaligned: the end effector is different, and the discrete steps a human takes to manipulate an object are not the same a two-finger gripper would need to take.

We never introduce this cross-embodiment gap in the first place. We hold morphology constant across every phase of learning by running the entire pipeline on hands. This is the bet we are making at mimic.

To balance our human data needs across qualities and quantities, we developed a strategy based on a “data pyramid”. At the base of the pyramid sits human video data: lower quality per sample, but available at massive scale. In the middle, wearable device data: higher quality than pure arbitrary egocentric video, but easier to scale than robot teleoperation. And only at the top, robot teleoperation and deployment data: the highest quality of data, collected directly with our own robots.

Infrastructure Built for High-performance Dexterity

A truly vertically integrated physical AI stack spans from hardware all the way to model inference. In between, a wide range of software components have to move in perfect harmony to unlock optimal performance. While it is tempting to rely on the research-focused solutions that carried the field so far, our bet is that exceptional performance and reliability in end-to-end robotics demand new, frontier infrastructure solutions.

What makes robotics infrastructure frontier? At their core, robotics software systems consist of real-time high-frequency control loops governing the low-level behavior of actuators and data acquisition from sensors. These are woven into a microservice architecture and are linked to each other and to external APIs by an inter-process communication (IPC) layer. Frontier infrastructure is one that makes no compromise on its IPC, keeps control performance high while minimizing its cost in compute, and meticulously exposes all of its inner state through interpretable, high-fidelity logging and telemetry. These principles guide the design of mimic’s in-house middleware stack that glues our hardware and AI policies together.

mimic-ipc: Lightning fast and reliable

During real-time inference of robot AI models, data rates easily reach Gb/s. Speed is not the only important aspect: end-to-end imitation learning performance is greatly dependent on low-jitter, synchronized sensor reading and actuation. Widely used robotics middleware options like ROS2 or LeRobot’s IPC implementation do not meet these tight requirements; both in speed (measured in message latency/data rates) and punctuality (measured in message jitter) they perform worse than current hardware bounds allow and bottleneck model performance. Our custom, hard real-time zero-copy middleware, mimic-ipc, brings a paradigm shift in both.

There is no technical reason for same-host communication to have a latency larger than a microsecond. Nevertheless, most current options waste CPU cycles where mimic-ipc does not make compromises. Comparing our custom solution to FastDDS with shared memory, we overall reach a ~18 000x speedup on HD image payloads.

Figure: Since they pay a high serialization/deserialization cost, the high latency of conventional middleware methods is especially observable on large image payloads. Comparing HD image payloads, mimic-ipc is a significant speedup at a 89 nanosecond median latency for processes running on separate cores.

While speed is also key in control loops, the real “end-to-end killer” is jitter. On a conventional ROS2 FastDDS setup, for example, this can reach milliseconds easily for HD image payloads. Even on a low-dimensional control topic with a 232-byte payload, mimic-pc has 350x less jitter, as it keeps jitters in the nanosecond domain.

A World-class Teleoperation Stack

Teleoperation, even with its inherent limitations outlined above, will remain important in last-mile data collection, assisting fleet response and ensuring production throughput in edge case failure scenarios. Having established the bedrock of our middleware, mimic-ipc, we built fast, performant control loops for all of our arms and hands on top of it. This does not only extend to the lower-level control loops in our hands and arms, but also to all parts of our teleoperation stack.

With its 15 active DoFs, the mimic hand M1 needs precise, low-latency, and high-bandwidth human-to-robot retargeting that is beyond the current state of the art. Returning to the roots of classical model-based path planning solutions, mimic’s in-house retargeter can leverage fast, sampling-based solutions that implicitly remove the noise artifacts of currently widespread gradient-based solutions. To raise the bar for the community both in the design and the benchmarking of retargeting algorithms, we have published a snapshot of our internal retargeter state of the art.

Spotless and Zero-cost Telemetry

A verbose and interpretable recording of the system’s inner state can mean the difference between a five-minute hotfix or a week of debugging, so the fastest control loop in the world is still worthless if it is not observable. There is a tradeoff as well: as important as it is, observability can also be a trap as much as a tool. Unnecessary logging layers can steal cycles for the control loop; unstable loggers can silently drop samples or corrupt measurements. Frontier telemetry therefore, while recording everything, has to distort nothing, and cost as close to zero in compute as physically possible.

mimic’s telemetry and data recording system is built exactly for this. It reuses mimic-ipc’s zero-copy primitives, while moving actual persistence (encoding, compression, and writing out to a range of supported formats) to asynchronous processes. Crucially, observability and data collection are not two systems competing for resources, but a single, unified view of the robot. This view is not a mere post-hoc debug tool: it provides a real-time snapshot of mimic’s full fleet, allowing us to track system health and hardware performance down to the last tendon, unlocking predictive maintenance and lifecycle management of all our components.

Conclusion

Our mimic hand M1, the wearable U1, and the infrastructure that connects them are three parts of one bet: keep the human hand as a fixed morphology spec across hardware, data collection, and model training, in order to build a true general-purpose robotics foundation model and deployment platform. Developing the entire full-stack system in-house allows us to control and co-design these components to be maximally coherent, operating at the frontier of end-to-end robotics capabilities.

This technical moat enables us to pursue the development of our unique AI bet on Video Action Models, following up on mimic-video, our recent pioneering research work. Ultimately, dexterity is one of the last barriers between AI and the physical world. Solving it unlocks entire categories of physical work that automation has never been able to reach.

We will soon release an update on what comes next.