PP-Tac: Paper Picking Using Tactile Feedback in Dexterous Robotic Hands
Pei Lin 1,2⋆ Yuzhe Huang 1,3⋆ Wanlin Li 1⋆ Jianpeng Ma 1 Chenxi Xiao 2† Ziyuan Jiao 1†, 1 Beijing Institute for General Artificial Intelligence 2 ShanghaiTech University 3 Beihang University, ⋆ equal contributors † corresponding authors · 2025
Read PP-Tac: Paper Picking Using Tactile Feedback in Dexterous Robotic Hands by Pei Lin 1,2⋆ Yuzhe Huang 1,3⋆ Wanlin Li 1⋆ Jianpeng Ma 1 Chenxi Xiao 2† Ziyuan Jiao 1†, 1 Beijing Institute for General Artificial Intelligence 2 ShanghaiTech University 3 Beihang University, ⋆ equal contributors † corresponding authors (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-16649.txt.
Introduction
Pei Lin and co-authors build a robot hand that picks up paper using touch sensing in 2025. Thin, floppy items like paper defeat most robot hands. Seeing them is hard, and grip plans are scarce. Their hand carries fine touch sensors on many fingers. Slips are sensed live and grip force is tuned to stop them.
Grip moves come from a learned policy trained on staged pinching motions. Tests covered paper of varied stock, thickness, and stiffness. The system succeeded in near nine of ten tries. The authors claim a first for touch based paper grasping with a nimble hand. Helper robots for daily tasks are the target use.
Catalog record
| CREATOR | Pei Lin 1,2⋆ Yuzhe Huang 1,3⋆ Wanlin Li 1⋆ Jianpeng Ma 1 Chenxi Xiao 2† Ziyuan Jiao 1†, 1 Beijing Institute for General Artificial Intelligence 2 ShanghaiTech University 3 Beihang University, ⋆ equal contributors † corresponding authors |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-16649
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PP-Tac: Paper Picking Using Tactile Feedback in Dexterous Robotic Hands Pei Lin 1,2⋆ Yuzhe Huang 1,3⋆ Wanlin Li 1⋆ Jianpeng Ma 1 Chenxi Xiao 2† Ziyuan Jiao 1†, 1 Beijing Institute for General Artificial Intelligence 2 ShanghaiTech University 3 Beihang University, ⋆ equal contributors † corresponding authors Abstract Robots are increasingly envisioned as human companions, assisting with everyday tasks that often involve manipulating deformable objects. Although recent advances in robotic hardware and embodied AI have expanded their capabilities, current systems still struggle with handling thin, flat, and deformable objects such as paper and fabric. This limitation arises from the lack of suitable perception techniques for robust state estimation under diverse object appearances, and the absence of planning techniques for generating appropriate grasp motions. To bridge these gaps, this paper introduces PP-Tac, a robotic system for picking up paper-like objects. PP-Tac features a multi-fingered robotic hand with high-resolution round-shaped tactile sensors R-Tac for omnidirectional tactile sensing. This hardware configuration enables real-time slip detection and online force control that mitigates slips. Furthermore, grasp motion generation is achieved through a trajectory synthesis pipeline, which first constructs a dataset of fingers’ pinching motions. Then, a diffusion-based policy is trained to control the hand-arm robotic system. Experiments demonstrate that PP-Tac can effectively grasp paper-like objects of varying material, thickness, and stiffness, achieving an overall success rate of 87.5%. To the best of our knowledge, this work is the first attempt to grasp paper-like objects using a tactile dexterous hand. Fig. 1 : Overview of PP-Tac. The system leverages tactile feedback from the proposed round-shaped sensor (R-Tac), integrated into a dexterous robotic hand, to grasp thin, deformable, paper-like objects. (a) Hand motions are generated by a diffusion-based policy inspired by human strategies, such as sliding and pinching. (b) The hardware setup includes a robotic arm, a dexterous hand, and four fingertip-mounted tactile sensors that simultaneously detect force and slip events. I Introduction Robots are increasingly popular as assistive agents in everyday life, particularly within household environments [ 36 ] . These robots are designed to perform various domestic tasks, often involving the grasp of thin, deformable objects such as paper and fabric [ 48 ] . For instance, clothes-folding tasks [ 25 ] require high dexterity and adaptability to accommodate variations in fabric size, texture, and stiffness, while document organization tasks [ 1 ] demand picking capabilities for diverse paper types and form factors. Beyond domestic settings, handling deformable objects is essential in industrial and logistical applications, such as fabricating fabrics [ 5 ] and packing objects using plastic bags and cardboard [ 12 ] . Despite their significance, picking up paper-like objects remains challenging in robotics [ 48 ] . In particular, the main challenges are three-fold: 1) Vision systems, commonly used for manipulation, struggle to perceive contact information during interactions with deformable objects due to limited sensing modalities and occlusion, resulting in an inaccurate environment model for motion planning [ 26 ] ; 2) These objects are often flat in shape, lacking salient features for contact points and thus hindering the synthesis of stable grasps [ 9 ] . 3) The appearance of such objects exhibits high variability, as their shape undergoes continuous and unpredictable deformation during manipulation. These dynamic shape variations significantly impair the generalizability of vision-based methods. In contrast, humans excel at picking up paper-like objects by leveraging coordinated multi-fingered motion and tactile sensing. As shown in Fig. 1 (a), the process typically begins with establishing contact with the fingers, followed by sliding motions to deform the material and to generate a contact point for the pinched grasp. Such motion is made possible by hand’s high dof , which enables establishing multiple contact points adaptively during sliding motion. During this process, tactile sensing is also crucial as it allows humans to perceive the object’s deformation and decide the appropriate forces. These real-time adjustments ensure the successful execution of the picking-up action. Inspired by human strategies, this paper introduces a robotics system coined PP-Tac : P aper-like object P icking using Tac tile feedback. The PP-Tac system comprises two key hardwares: A dexterous robotic hand, and the associated hemispherical and high-resolution vbts ( vbts ) R-Tac. The fingertip-mounted tactile sensors provide real-time contact feedback during grasping operations. Featuring a spherical sensing area and a high-frame-rate monochrome camera, this design enables faster response times and simpler calibration processes compared to conventional RGB-based tactile sensors. An illustration of the system is shown in Fig. 1 (b). In addition to the tactile sensor, this paper also presents A diffusion-based motion generation policy (PP-Tac policy) that imitates human picking-up skills. The proposed method first employs efficient trajectory optimization to generate expert data replicating human sliding and pinching motions. Second, generalizability to diverse flat objects were acheived by training a diffusion policy using these trajectories, leveraging proprioceptive data and tactile feedback for adaptive control of the dexterous robotic hand. Comprehensive real-world experiments were conducted to evaluate the PP-Tac system. PP-Tac achieved an overall success rate of 87.5% in grasping everyday thin and deformable paper-like objects, including plastic bags, paper bags, and silk towels on flat surfaces. Fig. 1 (a) illustrates examples of our arm-hand system successfully picking up paper-like objects. The PP-Tac also demonstrates significant adaptability in picking up paper-like objects on various uneven surfaces. Additionally, an ablation study further validated the contributions of each system component, highlighting the critical role of vbts feedback and motion generation policies in achieving robust performance. To the best of our knowledge, this work represents the first demonstration of deformable object picking using a dexterous hand equipped with vbts . Overall, our contributions include: 1) We present R-Tac, a novel spherical tactile sensor designed with ease of fabrication, calibration, and scalable deployment. To demonstrate its utility, we integrate R-Tac into each fingertip of a fully actuated dexterous robotic hand, enabling real-time contact feedback during manipulation tasks. 2) We propose a novel trajectory-optimization-based data generation framework. The proposed framework does not rely on tactile or physical simulation, which is computationally expensive, and is capable of achieving robust sim-to-real transfer. 3) We present the PP-Tac policy , a diffusion-based control strategy that utilizes only tactile and robot proprioceptive feedback for manipulating paper-like objects. This approach demonstrates robust generalization across diverse materials and surface properties. 4) We provide the implementation and systematic experiments of the proposed algorithms on a real robot system. Both hardware and code for PP-Tac system are released to support further research and community development. II Related Work II-A Deformable Objects Manipulation Deformable Object Manipulation (DOM) aims to handle soft objects that alter shape during interaction, which has been a long-standing challenging task in robotic research. Challenges mainly arised from uncertainties in perception and complex soft-body dynamics [ 15 , 3 , 29 ] . Early approaches solved such problem using visual perception for state estimation [ 48 , 35 ] , enabling tasks like rope-handling [ 31 , 35 ] , cloth-folding [ 40 , 25 ] and picking up paper with marker [ 6094742 ] . However, vision-based methods often fall short in solving real-world DOM problems due to varying object appearance, object’s physical property that are usually unknown in advance, visual occlusions [ 23 , 6 ] , and