Hemlock reads workers’ muscles to teach robots a lighter touch

A robot can watch someone pick up an egg without learning how hard to squeeze. Hemlock is trying to supply that missing information by recording electrical activity in workers’ muscles alongside first-person video, then turning those signals into training data for robotic hands.

The startup’s wearable, called the Bracelet, measures muscle activity as a person handles objects and tools. Hemlock combines those recordings with video to produce annotations about force and contact. Its intended customers are robotics companies and developers of models that control machines in the physical world.

That makes the launch a data story as much as a hardware story. A recording can show where a hand moves and what it touches. The amount of pressure behind that movement is harder to establish from the picture alone. Hemlock wants demonstrations of everyday work to contain both kinds of information.

The company reports encouraging results from two narrow tests. In one, it collected a few minutes of human demonstrations of an egg pick-and-place task and used them to fine-tune Pi0.5, comparing video-only training with video plus Bracelet data. Hemlock says the added signals improved the success rate tenfold.

In a second demonstration, involving a humanoid hand screwing in a lightbulb, the company reports a 2.7-fold improvement using a model it identifies as Gemini 4 Argon. These are company-reported comparisons on particular tasks, rather than evidence that the same improvement transfers to arbitrary factory or household work. The launch does not give the absolute success rates or trial counts needed to put those multipliers in fuller context.

Watch Hemlock’s original launch demonstration.

Its collection method is designed around people doing work they already know. Hemlock says it equips skilled workers with the Bracelet and video capture gear during their normal activities, recording how they use their fingers and manipulate tools. Its website shows examples of the resulting data, including hand tracking and separate contact and force readouts.

The approach also reflects a practical constraint: recording a demonstration should not make the demonstration harder to perform. In its announcement, Hemlock argues that tactile gloves can obscure the hands and that instrumenting an entire environment is difficult to scale. Its wearable instead takes measurements at the wrist. Whether that produces sufficiently accurate annotations across different people and tasks remains central to the usefulness of the resulting data.

Founder Julian Olschwang previously studied electrical and computer engineering at Carnegie Mellon and worked at UCLA’s Biomechatronics Laboratory, according to his public YC profile. Hemlock is part of Y Combinator’s Fall 2026 batch.

The company is now looking for teams working on dexterous manipulation and physical AI training. Its immediate pitch is concrete: add another signal to the human demonstrations those teams already need, then test whether the robot handles the object more reliably.