New paper · September 2026

Same demonstrations. Half the grip force.

Human-to-robot skill transfer across the embodiment gap: no shared sensor, no robot in the loop. A robot hand with no tactile sensors learns to hold a cup gently.

Read the paper on arXiv

1.20 Ngrip force with the force channel

2.55 Ngrip force without it

25 of 25success, with and without force

50 pre-registered rollouts on a real robot hand, 25 per policy, one session. Grip force is estimated from the hand's motor current, the same instrument for both policies. One-sided Mann-Whitney p < 0.0001.

Both policies hold the cup every time. The one without force squeezes more than twice as hard, because nothing it observes tells it when the grip is enough. Extra force buys no extra success; on fragile objects it is how things get crushed.

Watch 12 seconds

Vision onlyThe thumb lands on the cup. It never grasps.

Without forceGrasps, then squeezes at 2.55 N.

With forceGrasps and holds at 1.20 N.

Same demonstrations, same policy recipe, same robot. Only the input differs. Head-camera view.

How it works

1
Wear the glove and do the task. No robot in the room.

2
The force the fingers felt is transferred from the glove to the robot.

3
The robot has no tactile sensors. It still learns how hard to hold.

How the force gets from the glove to a robot with no sensors is in the paper.

The result

Ablation: vision only never grasped (0 of 15); with hand state 25 of 25 at 2.55 N; with fingertip force 25 of 25 at 1.20 N — Vision only never grasped. Hand state makes it grasp. Force makes it gentle. The vision-only arm ran in a separate lamp-matched session (0 of 15, against 3 of 3 for interleaved controls) and stopped at 15 of 25 planned runs on a pre-declared time limit. Force is compared within one session only. Metrics and decision rules were fixed before scoring.

The hardware

The sensing glove, back of the hand, with the five flex sensors highlighted and the wrist electronics — Five fingertip force sensors, five flex sensors, a wrist IMU, one ESP32-S3. One clock for everything, plus a head camera.

The robot hand on its arm, covered by a plain black glove, with the fixed camera at top right — Aero Hand Open, seven actuators, no tactile sensors, on an SO-101 arm under a fixed camera. The glove on the robot is a plain cover with no sensors in it.

73 demonstration sessions, 38,150 frames, recorded with no robot present.

Limits. One object, one position, a fixed arm path; the policy controls the hand only. Force values come from a motor-current estimate: good for comparing the two policies, approximate in absolute newtons, and saturating near 5 N.

GIFT: Glove-Inferred Force Transfer

Force-aware human-to-robot skill transfer from a wearable sensing glove to a robot hand without tactile sensors. Seven pages, five figures.

Read the paper on arXiv Working on tactile sensing or robot learning? Email me