The hard part begins after the grasp
A robot can pick up a screwdriver and still be unable to drive a screw. Functional tool use requires the hand to resist forces at the tool tip, maintain its orientation and prevent rotation or sliding as the task changes. ARTiS, a laboratory gripper described in a preprint submitted on September 3, addresses that distinction with an active palm that conforms around a tool and then stiffens under vacuum.
Researchers Roman Mykhailyshyn, Yukiyasu Domae and Kensuke Harada tested the prototype with nine screwdrivers, two hammers and two drills. Their results provide evidence that the palm can carry much of the load during selected screwdriver tasks and that combining it with movable fingers improves resistance to forces at the tip. The study does not establish reliability on a production line. Its strongest contribution is narrower: it tests a hardware design while tools are applying force, instead of treating successful pickup as the final result.
A palm that changes from soft to rigid
ARTiS has seven controlled degrees of freedom. Its three rigid fingers can change their arrangement, while compliant fin-ray fingertips adapt to the surface they contact. The central palm contains granules inside a latex membrane. Before vacuum is applied, the material can deform around a handle. Removing air packs the granules together, increasing friction and forming a stiffer support around the object.
This process is known as granular jamming. The same physical principle appears when loose particles in a flexible bag become difficult to rearrange after the surrounding pressure increases. In ARTiS, jamming lets the palm act as a temporary mold for handles that differ in diameter, contour and surface texture. The fingers provide positioning, stabilization and reorientation rather than bearing the entire load alone.
The project documentation reports a palm response time of about 0.7 seconds with a 400-kilopascal air supply and maximum air consumption of 20 liters per minute or less. Those specifications help define the laboratory system, but they do not reveal its lifetime, maintenance needs or behavior after prolonged exposure to dust, impacts and repeated pressure cycles. The hand was mounted on a UR5e collaborative robot arm at an angle to a worktable, making the reported results properties of the complete experimental arrangement rather than the gripper in isolation.
Force tests isolate the palm's contribution
The researchers repeated each holding and force-testing condition 30 times. In tests with two screwdrivers, ARTiS resisted 5.4 and 7.8 times more force at the tool tip than a LEAP Hand using a two-finger spherical grasp. Against a three-finger power-wrap configuration on the same research hand, ARTiS recorded advantages of 20.4 and 32.8 percent. These comparisons support the value of wrapping support around the handle, but they do not show that ARTiS is categorically better than dexterous hands across other objects or tasks.
LEAP Hand was created as a relatively low-cost, anthropomorphic platform for robot-learning research. It offers dexterous finger motion and therefore represents a useful experimental baseline, not a universal industrial standard. ARTiS and LEAP also embody different design priorities. One emphasizes a conforming palm that can lock around a tool, while the other emphasizes controllable fingers. The comparison is most informative as evidence about the selected grasp configurations.
An ablation analysis, in which parts of the system were disabled or varied, attributed more than 90 percent of the average holding contribution for the tested screwdrivers to the jamming palm. Torque measurements added another view of the design. The ninth screwdriver reached a peak tightening torque of 4.79 newton-meters, while the fourth reached 2.64 newton-meters. The variation indicates that handle shape, contact geometry and tool selection still matter even when the underlying gripping method is unchanged.
Useful demonstrations came with uneven results
Across the staged experiments, Table II reports 97 percent success for initial jamming and lifting and 68 percent for the complete grasp, reorientation or use, and release sequence. The authors also report an 85 percent aggregate across the three tested stages. Demonstrations included driving screws, hammering a nail, using a hammer to strike a screwdriver and levering apart composite fasteners. These examples show a broader range than static holding, although each remains a controlled laboratory demonstration.
The failures are technically revealing. The complete sequence produced no successes in 30 trials for screwdrivers 3, 5 and 8, while screwdriver 1 completed 27 of 30 trials. The paper links release failures for the smooth fifth and eighth screwdrivers to their limited texture and protrusions, which gave the palm little structure to retain. The more serious boundary appeared with a powered drill: the first drill recorded no successful operating trials. According to the paper, its vibration compacted the palm's granules, opened a gap at the contact surface and left the fingers unable to retain the running tool by themselves.
A second drill completed the reported trials, which the researchers associate with its lower vibration. That contrast does not mean the electric-tool problem has been solved. The authors suggest possibilities such as changing feed force or speed, stiffening the palm, increasing vacuum or adding side restraints. None of those remedies was validated in this study. An industrial tool can produce changing vibration, torque, heat and debris over long shifts, conditions not reproduced by a short set of demonstrations.
A small study points to a human benefit
The team also examined whether the gripper could support a person during a screw-removal task. Twelve participants each performed six trials with each of two methods. In the first, a participant held an air-conditioning component with one hand and operated a screwdriver with the other. In the second, ARTiS stabilized the component, freeing both hands for unscrewing rather than requiring one hand to hold the workpiece.
Including eight seconds needed to set up ARTiS, the mean time to remove three screws fell from 39.12 seconds to 33.23 seconds, a difference of 5.89 seconds between the group means. Separately, the paper reports 13.77 percent as the mean of the participants' individual percentage savings. The paired statistical tests found a difference within this experiment, although one participant was slightly slower with robotic assistance. This is not evidence of a factory-wide productivity gain. It is a bounded indication that secure robotic fixation can free both hands and reduce time spent manually stabilizing a part.
That division of labor has practical appeal. A robot does not need to replace a skilled worker to be useful. Holding a component or maintaining a tool's pose can remove an awkward supporting action while leaving judgment and fine control with the person. Whether ARTiS provides that benefit safely and consistently would depend on further ergonomic, reliability and task-specific studies.
Progress toward tool use, not production readiness
The researchers also lifted selected objects from the YCB set, a standardized collection created to make manipulation experiments more comparable across laboratories. Such testing broadens the evidence beyond the featured tools, but lifting benchmark objects is not equivalent to performing useful work with them. The study offers no independent replication, long-duration endurance test or comparison across the full range of industrial grippers.
The authors identify several next steps, including structural optimization, a metal prototype, real-to-simulation modeling and integration with robot-learning methods. The present device uses extensively 3D-printed components, which contributed to its size. A metal version might change stiffness, mass and durability, but it remains future work rather than a demonstrated upgrade.
ARTiS therefore marks practical hardware progress at the laboratory stage. Its experiments quantify why a conforming, stiffened palm can outperform finger-only support for certain screwdriver grasps, while the drill failure exposes a mechanism that must be addressed before demanding deployment. The result is neither a general-purpose robot hand nor proof of production readiness. It is a testable design argument: for forceful tool use, the palm can be an active load-bearing component rather than empty space between robotic fingers.