New Framework Lets Robotic Hands Recompute Grip Forces in Real Time

Grasping something firmly is one thing; keeping hold of it while it wiggles, slips, or gets bumped is another. Most dexterous-hand controllers compute a fixed distribution of contact forces at the start of a grasp and stick with it, an approach that can break down the moment the object shifts, a sensor model is slightly off, or an outside force pushes on the hand.

A team led by Sang Min Kim has proposed a framework that instead recalculates the desired force distribution continuously, as reported in a paper submitted to arXiv (cs.RO / eess.SY). Rather than depending on dedicated tactile sensors at every contact point, the system geometrically estimates where the hand is touching the object using a tracked model of the object combined with proprioception — the robot's internal sense of its own joint positions.

At each control cycle, the framework works out how contact forces should be spread across the whole hand, subject to three practical constraints: friction limits at each contact (so fingers don't slip), actuator limits (so the motors aren't asked to exceed their capability), and what the authors call an actuation-consistency constraint, inspired by classical whole-limb force analysis, which keeps the commanded forces achievable given the hand's actual joint torques.

This force-regulation layer is paired with a reactive reaching controller, so the system can do more than just hold on — it can initiate a grasp, maintain it as disturbances occur, and regrasp automatically if the object is knocked loose.

The team tested the approach in two settings. In gravity-free simulation experiments, the new method showed better grasp retention under controlled perturbations compared with two baseline strategies: a fixed force allocation computed in advance, and a fingertip-only execution that ignores contacts elsewhere on the hand. In real-world tests on a 27-degree-of-freedom robotic arm-hand system, the controller maintained grasps and recovered them when a human physically disturbed the held object, even as the contact points shifted across the whole hand during the interaction.

The result points toward dexterous hands that behave less like rigid grippers with pre-planned finger positions and more like adaptive systems that continuously renegotiate their grip in response to what's actually happening at the contact surface — without needing an array of tactile sensors embedded across the hand.

The paper, titled "Real-Time Force Regulation for Whole-Hand Dexterous Grasping," spans 9 pages and includes 10 figures, alongside a project page referenced in the submission for additional demonstrations.