The HANDi Hand V2: Enhancing a Multi-articulating Prosthetic Hand for Continual Machine Learning Research
This paper presents the redesigned HANDi Hand V2, an open-source, sensorized prosthetic platform that bridges the performance gap with commercial devices and demonstrates how machine-learned adaptive switching can reduce user control effort while enabling diverse grasp selection.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
For people who have lost a hand, modern prosthetic devices can look like marvels of engineering, capable of mimicking the complex movements of a biological limb. Yet, for many users, these devices remain frustratingly difficult to control. The core problem is not that the machines cannot move, but that the human mind must work too hard to tell them what to do. To pick up a cup, a user might have to manually cycle through a menu of different grip styles, a process that demands intense focus and often leads to fatigue. When the mental effort required to operate a device becomes too great, users often stop using it altogether. Researchers are now exploring a different path: instead of asking the human to manage the machine's complexity, they are teaching the machine to anticipate the human's needs. This approach relies on a concept called continual learning, where a device observes a user's actions over time and gradually learns to predict their intentions, effectively becoming a partner that handles the heavy lifting of decision-making.
At the University of Alberta, a team of researchers has built a new tool to test this idea, addressing a critical gap in the field. While expensive commercial prosthetics exist, they are often closed systems that researchers cannot easily modify or instrument with the sensors needed for advanced learning experiments. Conversely, many low-cost, 3D-printed hands available for research lack the durability and sensory richness required to survive the rigorous testing needed for machine learning. The team responded by redesigning their open-source "HANDi Hand," creating a second version that is both tougher and smarter. This new device, the HANDi Hand V2, is built to withstand the repeated cycles of trial and error that machine learning requires, while also gathering a wide array of data about how it moves and touches the world.
The physical redesign focused on solving the fragility that plagued the original version. The researchers replaced the old motors with a new type that can run for at least one hour without overheating, a benchmark chosen to match the length of typical laboratory learning sessions. They also redesigned the internal tendons that pull the fingers closed. Instead of using strings and springs that could tangle or break, the new design uses heavy-duty nylon zip ties. These ties are strong enough to handle the stress of repeated grasping and can be swapped out in less than five minutes if they ever wear out. To make the hand more sensitive, the team added a suite of sensors that measure the angle of every joint, the force at the fingertips, and even the temperature and electrical current inside the motors themselves. A camera mounted on the wrist provides a view of the objects the hand is reaching for, giving the computer a clear picture of the task at hand.
To prove that this new hardware was ready for real-world use, the researchers put it through a standardized test called the Anthropomorphic Hand Assessment Protocol. This test involves attempting to grasp twenty-six common household objects using ten different types of grips, from holding a large bottle to pinching a small coin. The HANDi Hand V2 succeeded in 79 percent of the tasks, a score that places it ahead of other 3D-printed research hands and narrows the performance gap with expensive commercial devices. The improvement was particularly noticeable in the ability to hold objects steady; the new motors allowed the hand to maintain a firm grip even when the flexible tendons tried to relax, a crucial feature for tasks like carrying a plate without dropping it.
With a robust platform in hand, the team then tested whether the device could learn to choose the right grip on its own. They set up an experiment where a participant had to move the hand through a sequence of three different grips: an open hand, a pinch, and a cylindrical hold. In the first condition, the user had to press a button repeatedly to cycle through the options until the correct one appeared, a process that took about 45 seconds to complete a full cycle. In the second condition, the system used a learning algorithm to predict which grip the user wanted based on the direction the hand was pointing. As the user repeated the task, the system learned the pattern and began to present the correct grip immediately. Within a few minutes, the time required to switch between grips dropped to about 33 seconds, and the number of times the user had to press the button decreased significantly.
The results suggest that shifting the burden of control from the user to the machine is not only possible but effective. By the end of the experiment, the system had learned to anticipate the user's intent so well that it could often skip the menu entirely, presenting the desired action with a single signal. The researchers noted that this change did more than just save time; it removed the physical hesitation users felt when cycling through options, allowing them to approach objects with confidence. While this study was conducted in a controlled setting with a limited number of grips, the findings indicate that a durable, sensor-rich platform like the HANDi Hand V2 can support the development of prosthetics that act as intelligent extensions of the human body. The work demonstrates that with the right hardware, machines can learn to be less reactive tools and more proactive partners, potentially reducing the frustration that leads many to abandon their devices.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.