Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ScissorBot: Learning Generalizable Scissor Skill for Paper Cutting via Simulation, Imitation, and Sim2Real

Jiangran Lyu, Yuxing Chen, Tao Du, Feng Zhu, Huiquan Liu, Yizhou Wang, He Wang

TL;DR

This work tackles generalizable paper cutting with scissors by integrating a dedicated paper-cutting simulator, imitation learning with action primitives, and sim2real transfer. It introduces PaperCutting-Sim to model interactive fracture and enables large-scale demonstrations via a privileged Oracle policy, which are distilled into a vision-based policy that operates on multi-frame point clouds. Deviation correction and visual artifact mimicry bridge the sim-to-real gap, enabling deployment on a Realman robot with a single hand to achieve near-human performance on Easy tasks and strong results on Middle and Hard patterns, across multiple materials. The approach advances contact-rich deformable manipulation and demonstrates substantial improvements over baselines, with practical implications for automated, precise paper cutting and related fine-manipulation tasks.

Abstract

This paper tackles the challenging robotic task of generalizable paper cutting using scissors. In this task, scissors attached to a robot arm are driven to accurately cut curves drawn on the paper, which is hung with the top edge fixed. Due to the frequent paper-scissor contact and consequent fracture, the paper features continual deformation and changing topology, which is diffult for accurate modeling. To ensure effective execution, we customize an action primitive sequence for imitation learning to constrain its action space, thus alleviating potential compounding errors. Finally, by integrating sim-to-real techniques to bridge the gap between simulation and reality, our policy can be effectively deployed on the real robot. Experimental results demonstrate that our method surpasses all baselines in both simulation and real-world benchmarks and achieves performance comparable to human operation with a single hand under the same conditions.

ScissorBot: Learning Generalizable Scissor Skill for Paper Cutting via Simulation, Imitation, and Sim2Real

TL;DR

This work tackles generalizable paper cutting with scissors by integrating a dedicated paper-cutting simulator, imitation learning with action primitives, and sim2real transfer. It introduces PaperCutting-Sim to model interactive fracture and enables large-scale demonstrations via a privileged Oracle policy, which are distilled into a vision-based policy that operates on multi-frame point clouds. Deviation correction and visual artifact mimicry bridge the sim-to-real gap, enabling deployment on a Realman robot with a single hand to achieve near-human performance on Easy tasks and strong results on Middle and Hard patterns, across multiple materials. The approach advances contact-rich deformable manipulation and demonstrates substantial improvements over baselines, with practical implications for automated, precise paper cutting and related fine-manipulation tasks.

Abstract

This paper tackles the challenging robotic task of generalizable paper cutting using scissors. In this task, scissors attached to a robot arm are driven to accurately cut curves drawn on the paper, which is hung with the top edge fixed. Due to the frequent paper-scissor contact and consequent fracture, the paper features continual deformation and changing topology, which is diffult for accurate modeling. To ensure effective execution, we customize an action primitive sequence for imitation learning to constrain its action space, thus alleviating potential compounding errors. Finally, by integrating sim-to-real techniques to bridge the gap between simulation and reality, our policy can be effectively deployed on the real robot. Experimental results demonstrate that our method surpasses all baselines in both simulation and real-world benchmarks and achieves performance comparable to human operation with a single hand under the same conditions.
Paper Structure (32 sections, 4 equations, 12 figures, 7 tables)

This paper contains 32 sections, 4 equations, 12 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deformable Object Manipulation
  4. Simulation Environments for Paper Cutting
  5. Imitation Learning
  6. Method
  7. Task Formulation and Method Overview
  8. PaperCutting-Sim
  9. Dynamic Model.
  10. Interactive Fracture.
  11. Demonstration Generation
  12. Vision-based Imitation Learning
  13. Sim-to-Real Transfer
  14. Hardware System Design
  15. Experiments
  16. ...and 17 more sections

Figures (12)

  • Figure 1: Robotic paper cutting with scissors.Left: The objective is to drive scissors to accurately cut curves drawn on the paper, which is hung with the top edge fixed. Middle: Our execution follows an action primitive sequence, namely Rotate, Close, Open, Push. The meticulous action, visualized as scissors before (orange) and after (green) each action, ensures accurate cutting in the real world. Right: During execution, large deformation of paper and severe occlusion between scissors and target curves occasionally occurs. Project Page: https://pku-epic.github.io/ScissorBot/
  • Figure 2: An overview of the learning system. The system first generates expert demonstrations in our built simulation which supports interactive fracture of the paper. These demonstrations are then used to train a vision-based imitation learning policy that inputs multi-frame point clouds (Blade point cloud is highlighted in green only for visualization) and outputs parameters of action primitive. Meanwhile, Deviation Correction and Visual Artifact Mimicry provide data augmentation to imitation learning which ensures a robust transfer from simulation to real world.
  • Figure 3: Interactive Fracture in our PaperCutting-Sim. (1): As the scissors close, the fracture occurs along the cutting direction. Intersection points (red star) can be computed from edge-edge detection and vertex-face detection between the cutting direction (orange dashed) and the paper mesh. (2): (a) The original paper mesh (blue triangles). (b) Intersection point (red star) and cutting direction (orange dashed). (c) According to the intersection points, new vertices are added on the existing edges and the endpoint is inserted inside the triangle. The new edges (green solid) are connected between the new inserted vertex and the opposite vertex in the triangle. (d) The edges between these newly added vertices are split into two pieces (black solid).
  • Figure 4: Visualization of Visual Artifact Mimicing (a) Perfect point cloud in simulation with scissors blade highlighted (green points) (b) Point cloud with our proposed visual artifact mimicry. (c) Point cloud captured in the real world with artifact.
  • Figure 5: Hardware System
  • ...and 7 more figures