Grasp2Twist: Learning Bimanual Dexterous Jar Opening by Reinforcement Learning

Bimanual dexterous jar opening with a single RL policy, transferred zero-shot to hardware

Mo Xu1, Yunfu Deng1, Jianuo Wang2, Josiah Hanna1†, Bilge Mutlu1†

1University of Wisconsin-Madison   2Independent Researcher   †Joint senior authors

NSF Engineering Research Centers NSF Engineering Research Centers NSF HAND ERC NSF HAND ERC University of Wisconsin-Madison
arXiv Video Code

One RL-trained policy grasps and twists open jars with a bimanual dexterous system.

One hand reconfigures finger-object contacts to enable sustained lid rotation while the other hand holds the jar.

Lightweight design choices enable complex bimanual dexterity.

Video

Task and Method

Physical system (left) and simulation (right).

We guide the fingers into a spatial arrangement during grasping with a continuous enclosure measure and signal the grasp-to-twist transition with a binary enclosure indicator.

We train the policy with a three-stage curriculum.

STAGE 1

Foundational skill acquisition

Learn grasp acquisition and the grasp-to-twist transition.

STAGE 2

Contact reconfiguration exploration

Explore finger contact changes for sustained twisting.

STAGE 3

Sim-to-real adaptation

Build sim-to-real robustness with domain randomization.

Finger Gaiting in the Real World

Twist
Release
Recontact
Twist again

Fingers release, reposition, and re-establish contact to continue twisting beyond the motion range of a fixed grasp.

Physical Evaluation on Household Containers

8/10
Blue Peanut
9/10
Yellow Mayo
6/10
Pink Glass
10/10
Green Vitamin
10/10
Brown Peanut
10/10
Instant Coffee

88%

53 / 60 trials

Six household containers

10 trials per container

Success: complete lid thread disengagement, verified by open-loop lid lifting by the twisting hand.