MIT Engineers Develop Ultrasound Wristband Capable of Tracking Internal Muscle Movements to Teach Humanoid Robots Hand Skills

Replicating the nuanced dexterity of the human hand has long been a core obstacle in the field of robotics, where traditional tracking methods often fail to capture subtle finger paths. Researchers at the Massachusetts Institute of Technology (MIT) have developed a wearable ultrasound wristband that monitors internal muscle and tendon activity to continuously reconstruct human hand positions.

The device combines a miniaturized transducer sticker with an artificial intelligence (AI) algorithm to map the physical mechanics of the wrist. By capturing internal structural shifts, the system provides a continuous stream of data that can be used to train humanoid robots in intricate physical tasks.

Human hand motion relies on a complex network of 34 muscles and more than 100 tendons and ligaments. The fingers and thumb alone account for 22 degrees of freedom, which represent the distinct ways joints can extend, flex, and angle during manual work.

The system uses high-frequency sound waves to look beneath the skin, creating real-time images of the wrist mechanics that control finger positions. Xuanhe Zhao, an MIT professor of mechanical engineering, explained that the tendons and muscles function like puppet strings, meaning a picture of these structures reveals the exact state of the hand.

The hardware consists of an ultrasound sticker the size of a smartwatch, paired with onboard electronics that are comparable in size to a mobile phone. A soft hydrogel layer secures the sensor to the skin, allowing for continuous, mobile image capture without restricting natural hand mobility.

To translate the black-and-white ultrasound images into precise spatial coordinates, the researchers trained a specialized machine-learning algorithm. The AI model identifies specific patterns within the moving wrist tissue and correlates them with the corresponding degrees of freedom in the hand.

During laboratory testing, eight volunteers with varying wrist and hand sizes evaluated the system. The device tracked smooth transitions between complex gestures, including all 26 letters of American Sign Language, and monitored the manipulation of everyday items like scissors, pencils, and plastic bottles.

The wristband operates with a response latency of within 120 milliseconds and functions wirelessly. This wireless capability allows an operator to guide a mechanical hand remotely, even if the receiving robot is positioned in a separate room.

In live demonstrations, users successfully commanded a commercial robotic hand to replicate physical movements. The system allowed a person wearing the wristband to manipulate the robot into playing a simple tune on a keyboard and executing taps in a desktop basketball game.

The interface also translates hand motions into virtual environments without requiring external cameras. Wearers were able to grab, rotate, and manipulate digital objects on a computer screen, such as pinching fingers together to alter the size of an asset.

While teleoperation provides immediate utility, the research team is focusing on data collection to build a large-scale repository of human motion. This dataset aims to support training pipelines that would allow humanoid platforms to learn delicate manual execution without direct human supervision.

The wearable imaging approach avoids several limitations inherent to alternative tracking systems. Camera-based setups are prone to visual occlusion and environmental noise, while sensorized gloves cover the hand and eliminate natural tactile feedback during object interaction.

Other tracking methods rely on surface electrical signals from muscles, but these inputs are highly sensitive to environmental interference. Electrical sensors can often detect broad actions, such as an open or closed fist, but struggle to map the intermediate paths between movements.

The research, which was detailed in the journal Nature Electronics, received institutional backing from several organizations. Funding was provided by MIT, the U.S. National Institutes of Health (NIH), the U.S. National Science Foundation (NSF), the U.S. Department of Defense (DoD), and the Singapore-MIT Alliance for Research and Technology (SMART).

Future engineering work will focus on further miniaturization of the wristband hardware. The developers also intend to expand the underlying AI model to ensure compatibility with a wider diversity of user anatomies and behavioral gestures without requiring long personal training sessions.