A robotic hand adjusted its grip as an object was pushed and twisted, using a tracked object shape and the hand’s joint positions instead of touch sensors at its contact points. Researchers at MIT, Seoul National University and Yonsei University report that the controller regrasped objects after six of eight slips in a small hardware demonstration.
A planned grip can go stale as the object moves: fingers and palm meet it at new points, so forces assigned to the original contacts may no longer stabilize it. The team’s controller estimates contacts across the hand’s fingertips, finger sides and palm from the object’s tracked surface and the hand’s joint readings. It then reallocates forces as those contacts change, subject to friction and motor-torque limits.
The approach estimates contact points from the object’s shape, but it still depends on seeing and modeling the object. In simulation, researchers tested 50 grasps per method across five representative objects, with ten trials per grasp at rotation targets from 30 to 90 degrees, including initial failures. At 90 degrees, the whole-hand method succeeded 88.6% of the time, compared with 72.4% for a fingertip-only version. Gravity was disabled, so the simulation did not test the effect of weight.
The physical setup paired a seven-degree-of-freedom arm with a 20-degree-of-freedom, five-finger hand. An RGB-D camera tracked the object at 30 frames per second; the hand’s fingertip tactile sensors were not used. Across three runs involving a can and two boxes, people applied 78 pushes and twists over about 200 seconds. The hand lost the object eight times and independently recovered six; in the other two cases, the object moved outside the arm’s workspace.
The system still needs an object model and reliable object tracking. Occlusion can spoil contact estimates, and small tracking errors are especially difficult for thin objects. Its force calculation does not explicitly account for gravity, inertia or loads that change during movement; the authors list degraded tracking and dynamic effects as open issues.
Sources
- Kim et al., “Real-Time Force Regulation for Whole-Hand Dexterous Grasping,” arXiv:2609.30082v1 (24 September 2026).
- Authors’ project page.
- Illustrative image: “Robot hand,” by lost places via Wikimedia Commons, CC BY 2.0. The photograph does not show the study hardware.
.jpg){kind=link}