Robots Can See, But CMU's New Glove Teaches Them to Feel

Robots are getting better at manipulating objects, but they still struggle with many tasks that are easy for humans. If these machines are going to be of much use in our world, they need to learn how to behave more like us. The problem is that we don’t have a great way to teach them how to do this.

Cameras can track motion, and motion-capture gloves can measure finger positions, but neither tells the whole story. When you pick up a screwdriver, turn a key, or open a jar, success depends not only on where your fingers are, but also on where and how they make contact with the object. Researchers at Carnegie Mellon University have developed a new glove designed to capture exactly that missing information.

Called ART-Glove, short for Articulated Tactile Glove, the system combines joint tracking with high-resolution tactile sensing. The goal is to record human demonstrations of dexterous manipulation in a way that can later be used to train robotic systems.

One of the challenges in collecting training data for robot learning is balancing measurement accuracy with natural movement. Existing approaches often force a compromise. Teleoperation systems can produce robot-ready data, but they typically provide limited tactile feedback and can feel unnatural to operate. Vision-based systems preserve natural hand motion, but important details about contact are hidden from view or obscured during manipulation. Soft sensing gloves add data collection capabilities, yet their flexible surfaces make it difficult to accurately determine contact geometry.

ART-Glove takes a different approach. Instead of treating the hand as a collection of invisible joints, it represents it using 16 rigid functional surfaces distributed across the fingers, thumb, and palm. These surfaces define exactly where contact can occur. Each finger is divided into distal, middle, and proximal sections, while the thumb receives three dedicated segments and the remaining palm area is covered by a rigid shell. Together, they create a known contact geometry that can be tracked during manipulation.

To ensure these rigid surfaces move naturally with the wearer’s hand, the researchers connected them using 22 anatomically aligned joints. The design tracks finger flexion, extension, abduction, and other motions while attempting to preserve normal dexterity. Rather than relying on bulky external mechanisms, the glove uses a mix of specialized joint designs that fit within the limited space available around the hand.

Every functional surface is covered by a piezoresistive tactile skin containing a total of 2,048 sensing elements, or taxels. These sensors measure contact pressure across the glove’s surface. Joint motion is captured by magnetic rotary encoders positioned throughout the mechanism. An STM32F411 microcontroller gathers data from both systems and synchronizes the measurements at 120 Hz. This produces a stream of 22 degrees-of-freedom worth of joint data paired with detailed tactile information collected in real time.

For roboticists, that combination could prove valuable. A robot learning from human demonstrations needs more than a record of hand motion. It also needs to understand which surfaces touched an object, when contact occurred, and how those interactions changed throughout a task. ART-Glove was designed to capture all of that information in a single wearable device.