Uncertainty-Aware Finger Force Estimation for Reliable Box Manipulation

Anonymous Authors

TL;DR FEEL aims to enable reliable box pick-and-place through per-finger force estimation from proprioception and adaptation, without dedicated contact sensors during deployment.

01 / BOX PICK-AND-PLACE

What changes when fingers adapt to force?

TL;DR FEEL uses estimated force to adapt each finger and places the box intact; fixed commands keep closing after contact and damage the seam in the right-hand clip.

With force adaptation

FEEL

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

Without force adaptation

Baseline

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

Both conditions used the same box specification and predefined arm motion. Reliability requires both task success and a damage-free outcome.

Reliability

Primary status for all experiment grids below.

Reliable Unreliable

Reliable indicates task success and a damage-free outcome. Each card reports both criteria separately.

With force estimation experiments

K_p = 200

TL;DR At the 6-N threshold, FEEL holds or opens contacted fingers while others continue closing, enabling the box transfer without seam damage in all ten trials.

Trial 01Reliable

Task success: Yes
Damage-free: Yes

Trial 02Reliable

Task success: Yes
Damage-free: Yes

Trial 03Reliable

Task success: Yes
Damage-free: Yes

Trial 04Reliable

Task success: Yes
Damage-free: Yes

Trial 05Reliable

Task success: Yes
Damage-free: Yes

Trial 06Reliable

Task success: Yes
Damage-free: Yes

Trial 07Reliable

Task success: Yes
Damage-free: Yes

Trial 08Reliable

Task success: Yes
Damage-free: Yes

Trial 09Reliable

Task success: Yes
Damage-free: Yes

Trial 10Reliable

Task success: Yes
Damage-free: Yes

Without force estimation experiments

K_p = 200

TL;DR Without force adaptation, fixed finger commands continue closing after contact. Placement still succeeds, but the box seam is damaged in three trials, so task success alone is not reliable handling.

Trial 01Unreliable

Task success: Yes
Damage-free: No

Trial 02Reliable

Task success: Yes
Damage-free: Yes

Trial 03Reliable

Task success: Yes
Damage-free: Yes

Trial 04Reliable

Task success: Yes
Damage-free: Yes

Trial 05Reliable

Task success: Yes
Damage-free: Yes

Trial 06Reliable

Task success: Yes
Damage-free: Yes

Trial 07Reliable

Task success: Yes
Damage-free: Yes

Trial 08Reliable

Task success: Yes
Damage-free: Yes

Trial 09Unreliable

Task success: Yes
Damage-free: No

Trial 10Unreliable

Task success: Yes
Damage-free: No

02 / CONTROLLER SENSITIVITY

Kp comparison

TL;DR Kp = 100 estimates force most accurately but often fails the transfer; Kp = 300 makes fingers too stiff for stable grasp. Kp = 200 balances both for reliable, damage-free placement.

K_p = 1004/10 reliable

Trial 01 · Task not completed

K_p = 20010/10 reliable

Trial 01 · Task completed, box intact

K_p = 3007/10 reliable

Trial 09 · Task completed, box damaged

K_p = 100

RMSE: 0.233 ± 0.005 N

10 trials

Trial 01Unreliable

Task success: No
Damage-free: Yes

Trial 02Reliable

Task success: Yes
Damage-free: Yes

Trial 03Unreliable

Task success: No
Damage-free: Yes

Trial 04Unreliable

Task success: No
Damage-free: Yes

Trial 05Unreliable

Task success: No
Damage-free: Yes

Trial 06Unreliable

Task success: No
Damage-free: Yes

Trial 07Reliable

Task success: Yes
Damage-free: Yes

Trial 08Reliable

Task success: Yes
Damage-free: Yes

Trial 09Reliable

Task success: Yes
Damage-free: Yes

Trial 10Unreliable

Task success: No
Damage-free: Yes

K_p = 200

Used in the main experiments

RMSE: 0.251 ± 0.010 N

10 trials

Trial 01Reliable

Task success: Yes
Damage-free: Yes

Trial 02Reliable

Task success: Yes
Damage-free: Yes

Trial 03Reliable

Task success: Yes
Damage-free: Yes

Trial 04Reliable

Task success: Yes
Damage-free: Yes

Trial 05Reliable

Task success: Yes
Damage-free: Yes

Trial 06Reliable

Task success: Yes
Damage-free: Yes

Trial 07Reliable

Task success: Yes
Damage-free: Yes

Trial 08Reliable

Task success: Yes
Damage-free: Yes

Trial 09Reliable

Task success: Yes
Damage-free: Yes

Trial 10Reliable

Task success: Yes
Damage-free: Yes

K_p = 300

RMSE: 0.255 ± 0.008 N

10 trials

Trial 01Reliable

Task success: Yes
Damage-free: Yes

Trial 02Reliable

Task success: Yes
Damage-free: Yes

Trial 03Reliable

Task success: Yes
Damage-free: Yes

Trial 04Unreliable

Task success: No
Damage-free: Yes

Trial 05Reliable

Task success: Yes
Damage-free: Yes

Trial 06Unreliable

Task success: Yes
Damage-free: No

Trial 07Reliable

Task success: Yes
Damage-free: Yes

Trial 08Reliable

Task success: Yes
Damage-free: Yes

Trial 09Unreliable

Task success: Yes
Damage-free: No

Trial 10Reliable

Task success: Yes
Damage-free: Yes

03 / REAL-ROBOT SUPERVISION

Data collection

TL;DR Across ten fingers, a wearable F/T sensor labels varied contacts while synchronized proprioceptive histories capture each response; these real-robot data enable force feedback without contact sensors during box manipulation.

L Thumb

L Index

L Middle

L Ring

L Little

R Thumb

R Index

R Middle

R Ring

R Little

Approximately 25,000 synchronized estimator samples and an additional 5,000 calibration samples were collected per finger. Robot states and F/T measurements were aligned by timestamp.

04 / PROPRIOCEPTIVE AMBIGUITY

Predictive uncertainty

TL;DR Force through structure varies load with little change in proprioception, unlike Force through joints; higher epistemic uncertainty flags this unresolved force ambiguity rather than recovering the missing information.

Force through joints

Standard finger-motion contact · epistemic uncertainty 0.102 ± 0.005

Force through structure

Load applied while the commanded posture stays fixed · epistemic uncertainty 0.172 ± 0.088

Component disagreement is an empirical signal of ambiguity, not a way to recover force information absent from proprioception.