Research

ergoCub Won’t Drop the Box Until You Mean It

Robb Harlan 5 min read

A humanoid holding a package has to know the difference between a real take and a bump. Pasquale Marra, Stefano Berti, Gabriele M. Caddeo, and Lorenzo Natale at the Italian Institute of Technology (with the University of Genoa) fine-tune NVIDIA Isaac GR00T N1.5-3B on that problem. The paper, arXiv:2609.05282, went up 4 September 2026.

The robot is ergoCub. Both hands wear Xela fingertip taxels. The prompt is “Get the box and pass it to human.”

ergoCub handing a brown box to a person, with force overlays in simulation
Release when the pull is toward the person; hold when the force is wrong-way. Source: arXiv:2609.05282.

Tactile history plus a soft arm

Vision alone cannot tell a sustained pull from a tap. The policy runs at 10 Hz; the taxels run near 70 Hz. The authors pack about 0.9 s of three-axis fingertip forces into a 24-dimensional Temporal Tactile Encoding (TTE) instead of a single-frame force vector.

A compliance layer yields the hands along measured fingertip forces (Kp=50, Kd=40) before the torso-and-arms inverse-kinematics controller. GR00T still outputs 16-step Cartesian chunks; they execute the first 8.

Eight people gave 50 teleop demos each (25 with compliance, 25 without) on a Meta Quest 3. Then 10 new participants ran 10 trials per policy: five release-expected, five hold-expected. Three policies:

  • P1: TTE + compliance
  • P2: compliance, no tactile
  • P3: TTE, no compliance
Sequence of a person pulling a box from ergoCub, brief pulls vs a sustained pull
Same peak force (~3.8 N); only the long pull opens the hands. Source: arXiv:2609.05282.

The scores

Over 300 scored trials:

P1 (full)P2 (no tactile)P3 (no compliance)
Success93%54%82%
Success with 2.8 s release window92%45%80%
Median release delay1.70 s2.35 s2.00 s
Peak pull force4.07 N4.67 N4.74 N
Satisfaction (1–7)6.23.34.7
Ranked first / safest9 / 101 / 00 / 0

P3 dropped the box in 4% of its trials — the only policy that did. Instantaneous force (an earlier ablation) false-opened 51 times on a held-out “challenging” set; TTE did it 3 times and never held a false opening longer than 1.5 s.

T9, short impulsive pulls at genuine-take magnitudes, is the hard case (P1 at 60%). Code and data are promised after acceptance.

A Human’s Take

Release timing is the whole job. Tactile history decides; compliance makes the hand-off feel less like a tug-of-war. 93% with naive users is the number I will remember. I still want this on an object that is not a cardboard box, and I want the thumbs in the sensor set. Four fingers per hand is a start.

Sources