Hemlock

Companies building physical AI and robotic systems need to teach machines how skilled humans use their hands, including the force and contact involved in manipulation tasks. Standard video datasets capture visual motion but not the force or contact information that is critical for a robot to replicate delicate or skilled hand movements. Without this data, physical AI models have an incomplete picture of how tasks are physically performed.

Hemlock operates as a data foundry for physical AI, producing egocentric video paired with contact and force annotations. The data is captured using the Bracelet, Hemlock's proprietary EMG wearable, which appears to record force and muscle activity while a skilled worker performs a task. The site presents case studies pairing this data with named AI models or systems (π0.5, Gemini, Argon), and shows sample data with overlays such as bones, mesh, contact points, wrists, and depth. The company positions this as 'muscle memory for machines,' letting robots learn skilled hand use from human demonstration.