Uncertainty-Aware Finger Force Estimation for Reliable Box Manipulation

Anonymous Authors

Real-robot experimentsForce-adaptive box pick-and-place in the experimental workspace.

With and without force adaptation

Without force adaptation

Baseline

Task success 10/10 · Damage-free 7/10 · Reliability 7/10

With force adaptation

FEEL

Task success 10/10 · Damage-free 10/10 · Reliability 10/10

10 trials per condition, using the same box specification and predefined motion. Reliability = successful placement without recorded damage.

Inside the experiments

Boxes used in the manipulation experiments.
Example of the recorded seam damage.
Example of a damage-free outcome.

Box dimensions: 395 × 275 × 240 mm · Mass: 2.0 kg · Damage criterion: tape separation at the designated seam.

Method overview

A calibrated virtual spring and residual MDN provide per-finger force feedback. Dedicated contact sensors are not used during manipulation.

Grasping across controller gains

(100, 200)

PD gains
4/10 reliabilityLowest estimation RMSE: 0.233 N.

(200, 400)

Selected setting
10/10 reliabilityHighest reliability among the tested settings.

(300, 600)

PD gains
7/10 reliability9/10 task success; 8/10 damage-free outcomes.

Representative trials at each gain setting. Reported rates summarize 10 trials per setting.