A Soft Wrist with Anisotropic and Selectable Stiffness for Robust Robot Learning in Contact-rich Manipulation

Steven Oh; Tomoya Takahashi; Cristian C. Beltran-Hernandez; Yuki Kuroda; Masashi Hamaya

cs.ROFeb 16, 2026

A Soft Wrist with Anisotropic and Selectable Stiffness for Robust Robot Learning in Contact-rich Manipulation

Authors: Steven Oh, Tomoya Takahashi, Cristian C. Beltran-Hernandez, Yuki Kuroda, Masashi Hamaya

Organizations: OMRON SINIC X Corp., Bunkyo-ku, Tokyo, 113-0033, Japan.

Abstract

Contact-rich manipulation tasks in unstructured environments pose significant robustness challenges for robot learning, where unexpected collisions can cause damage and hinder policy acquisition. Existing soft end-effectors face fundamental limitations: they either provide a limited deformation range, lack directional stiffness control, or require complex actuation systems that compromise practicality. This study introduces CLAW (Compliant Leaf-spring Anisotropic soft Wrist), a novel soft wrist mechanism that addresses these limitations through a simple yet effective design using two orthogonal leaf springs and rotary joints with a locking mechanism. CLAW provides large 6-degree-of-freedom deformation (40mm lateral, 20mm vertical), anisotropic stiffness that is tunable across three distinct modes, while maintaining lightweight construction (330g) at low cost ($550). Experimental evaluations using imitation learning demonstrate that CLAW achieves 76% success rate in benchmark peg-insertion tasks, outperforming both the Fin Ray gripper (43%) and rigid gripper alternatives (36%). CLAW successfully handles diverse contact-rich scenarios, including precision assembly with tight tolerances and delicate object manipulation, demonstrating its potential to enable robust robot learning in contact-rich domains. Project page: https://project-page-manager.github.io/CLAW/

Figures & tables

Fig. 1: The CLAW soft wrist: (a) A novel soft wrist featuring two orthogonal leaf springs to achieve anisotropic stiffness and enabling 6-DoF compliance. (b) CLAW’s extensive evaluation across a range of challenging contact-rich manipulation tasks using imitation learning.

Specification		Value			Unit
Specification		Ours 6-DoF	Wrist [ 18 ] 6-DoF	Wrist [ 19 ] 2-DoF	Unit
Weight		330 (w/ gripper, wrist, camera)	Not reported	110 (w/o gripper)	g
Size		155 × 150 × 85 (W × H × L)	$\phi$ 75 × 60	$\phi$ 63 × 45	mm
Deformation	X/Y	40	6.35	N/A	mm
	Z	20 (comp.), 10 (ext.)	11	N/A	mm
	Roll/ Pitch	15	8	20	Deg

TABLE I: Mechanical characterization of CLAW and other soft wrists

Fig. 2: Mechanical design of the CLAW: Orthogonal leaf springs and rotary joints provide 6-DoF compliance with anisotropic stiffness. A simple locking mechanism enables mode-switchable stiffness while preserving low weight and structural simplicity.

Fig. 3: Deformation characteristics of the CLAW compliant module: Two orthogonal looped leaf springs with rotary joints and coil spring preload enable 6-DoF anisotropic compliance.

Fig. 4: Length and range of motion of the leaf spring: (a) Fixed section of the leaf spring and dimensions of each part, (b) Maximum deformation state in the XY direction.

Fig. 5: Locking mechanism of the CLAW: A single actuator moves four pins with different profiles on a carrier to engage grooves in joint links. This system enables three stiffness modes: (a) Free mode with all joints rotatable, (b) Half-lock mode restricting motion along the X-axis by locking two joints, and (c) Full-lock mode locking all joints to increase stiffness in X, Y, and yaw while retaining compliance in other directions.

Fig. 6: Teleoperation setup: A haptic teleoperation leader device is employed to control a UR5e robot arm. On the leader device, user can select stiffness mode by switching a lever. FT sensor on the follower arm relay external force information to the leader, which is then mapped to joint torques to create haptic feedback.

Fig. 7: Deformation comparison of CLAW, Fin Ray, and a rigid gripper: . Each end-effector’s tool center point was fixed to a vertical pole, as shown on the left side. For the Z-axis deformation on the Fin Ray, we placed the gripper on top of the vertical pole. Force readings were measured via the UR5e’s onboard FT sensor.

Fig. 8: Door opening test: Top figures show the motivation for variable stiffness in door opening as humans turn doorknobs with high wrist stiffness and release with low wrist stiffness. The bottom figure shows a comparison of the force trajectories of variable stiffness, free mode only, and full-lock only. Free mode triggered the emergency stop as the force threshold was exceeded in a different axis.

Fig. 9: Learning performance benchmarking: (a) Experimental setup with the FMB board and insertion peg. (b) Representative trajectory from the rollout insertion path of the three end-effectors. (c) Success rate comparison between the learning performance of CLAW, Fin Ray, and rigid gripper for different peg geometries and approach directions.

Task	Success Rate	Stiffness Mode
Vertical peg-in-hole	0.90	Free
Glass-tube insertion	0.85	Free
Dish-cap placement	0.68	Free
Drawing smiley face	0.47	Free
Dragging metal peg	1.00	Half-lock
Horizontal peg-in-hole	0.45	Full-lock

TABLE II: Task success rates for various real-life tasks.

Soft Robotics

Soft robotics studies robots with compliant bodies and actuators that deform to move, sense, and interact with uncertain environments.

See all 231 papers

Beyond the End-Effector: Learning Whole-Arm Contact Manipulation from Human Demonstrations
Oct 13, 2026Bo-Ruei Huang, Daniel Seita, Erdem BıyıkNon-Prehensile Robotic Manipulation Robotic Manipulation
Thomas Lord Department of Computer Science, University of Southern California
Functional Force-Aware Retargeting from Virtual Human Demos to Soft Robot Policies
Apr 1, 2026Uksang Yoo, Mengjia Zhu, Evan Pezent +8Motion Retargeting Robot Imitation Learning
The Robotics Institute, Carnegie Mellon University · Meta Reality Labs
Tactile Memory with Soft Robot: Robust Object Insertion via Masked Encoding and Soft Wrist
Jan 27, 2026Tatsuya Kamijo, Mai Nishimura, Nodoka Shibasaki +3Robotic Control Dexterous Manipulation
OMRON SINIC X Corporation, 5-24-5 Hongo, Bunkyo-ku, Tokyo, 113-0033, Japan · The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-8654, Japan · Northeastern University, 360 Huntington Ave, Boston, MA, 02115, USA

Oct 13, 2026J/K move · Enter open · S save

Beyond the End-Effector: Learning Whole-Arm Contact Manipulation from Human Demonstrations

Bo-Ruei Huang, Daniel Seita, Erdem Bıyık

Thomas Lord Department of Computer Science, University of Southern California

Non-Prehensile Robotic Manipulation Robotic Manipulation Contact-Rich Robotic Manipulation Robot Imitation Learning Whole-Body Robotic Manipulation

Open paper AskPDF arXiv

Robot manipulation has traditionally relied on end-effector-centric interactions. In contrast, humans routinely exploit whole-arm contacts to improve efficiency, robustness, and adaptability, especially when the hands are occupied or when precise grasping is unnecessary. Enabling similar capabilities in robots remains challenging: the robot must select where on its arm to make contact, generate effective non-prehensile interactions, and satisfy task-specific pose or safety constraints on the end-effector. We propose Whole-Arm Contact Learning (WACL), a hierarchical framework that describes each whole-arm contact task with three robot-agnostic representations and hands them to a quasi-static surface-contact optimizer: semantic keypoints that summarize the scene, a reference contact trajectory that specifies where and how the arm presses on the object, and safety-aware path constraints that a vision-language model generates from the task description to protect the end-effector's task. The robot's kinematic chain enters only in this optimizer, which grounds these representations onto the arm by jointly searching over which link makes contact, where on that link, and the joint configuration that realizes it. This decomposition lets WACL learn the contact trajectory from 10 bare-hand human demonstrations per task collected without teleoperation, while new end-effector constraints change only the optimizer's input rather than any trained component. Real-world experiments on a 7-DoF arm across four whole-arm tasks show that WACL sustains contact under end-effector constraints where a strong link-pose baseline fails (78% vs. 33% success) and generalizes to new constraints without retraining. Website: https://wacl-robot.github.io

Abstract

Soft Robotics

Explore similar work

Beyond the End-Effector: Learning Whole-Arm Contact Manipulation from Human Demonstrations

Functional Force-Aware Retargeting from Virtual Human Demos to Soft Robot Policies

Tactile Memory with Soft Robot: Robust Object Insertion via Masked Encoding and Soft Wrist