Dex Hand

A robotic hand is only as good as the system that controls it. Foundation models gave robots a brain; mobile bases gave them a body. What’s missing is the control layer that lets a hand actually do the work, such as picking a strawberry, threading a cable, or handing you a mug by the handle, while also being capable in use cases that involve higher loads.

We’re building hardware with AI control in mind. Each phase of the hand is designed to teach the next: a 3D-printed prototype to prove the mechanism, a more refined build with a custom MCU board built-in, and a production-ready design tuned for the learned policies it has to run.

Mechanics, firmware, and reinforcement learning, co-designed as one system rather than three, with cost and manufacturing as primary constraints. The goal is to deliver a high-quality hand with advanced AI control.

Our hand uses reinforcement learning to perform dexterous manipulation tasks. An advanced algorithm is used to train the hand until it demonstrates the intended action. The model inputs will include positional feedback, fingertip force feedback, and 6-degree force and torque of the wrist.

By utilizing sensor-fusion with the model, the hand can perform tasks while operating under highly dynamic conditions. By allowing the model to interact with the real world, the system can produce a more informed output. This output, when paired with the robust mechanical design of the hand, allows for a wide range of use cases.