DiSECt: A Differentiable Simulation Engine for Autonomous Robotic Cutting

TL;DR

DiSECt introduces the first differentiable simulator for cutting soft materials by combining FEM with a continuous contact model based on signed distance fields and a continuous damage model realized with cutting-plane springs. This differentiability enables gradient-based parameter inference, faster calibration against ground-truth data, and efficient trajectory optimization for cutting actions. The approach demonstrates accurate calibration to commercial solvers and real-world measurements, robust generalization across velocities and geometries, and practical control insights such as pressing-and-sawing strategies. These capabilities position differentiable cutting simulators as valuable tools for rapid prototyping, sim-to-real transfer, and planning in robotics and potentially surgical applications.

Abstract

Robotic cutting of soft materials is critical for applications such as food processing, household automation, and surgical manipulation. As in other areas of robotics, simulators can facilitate controller verification, policy learning, and dataset generation. Moreover, differentiable simulators can enable gradient-based optimization, which is invaluable for calibrating simulation parameters and optimizing controllers. In this work, we present DiSECt: the first differentiable simulator for cutting soft materials. The simulator augments the finite element method (FEM) with a continuous contact model based on signed distance fields (SDF), as well as a continuous damage model that inserts springs on opposite sides of the cutting plane and allows them to weaken until zero stiffness, enabling crack formation. Through various experiments, we evaluate the performance of the simulator. We first show that the simulator can be calibrated to match resultant forces and deformation fields from a state-of-the-art commercial solver and real-world cutting datasets, with generality across cutting velocities and object instances. We then show that Bayesian inference can be performed efficiently by leveraging the differentiability of the simulator, estimating posteriors over hundreds of parameters in a fraction of the time of derivative-free methods. Finally, we illustrate that control parameters in the simulation can be optimized to minimize cutting forces via lateral slicing motions. We publish videos and additional results on our project website at https://diff-cutting-sim.github.io.

Figures (21)

• Figure 1: A rendering from our differentiable cutting simulator. DiSECt provides accurate gradients of the cutting process, allowing us to efficiently fit model parameters to real-world measurements, and optimize cutting motions.
• Figure 2: Visualization of an apple slice. We use a tetrahedral FEM-based model of the apple generated from scanned real-world data jamdagni2019robotic.
• Figure 3: To smoothly model damage, we insert springs (shown in blue) between cut elements. We incrementally reduce the spring stiffness based on damage over time, proportional to the contact force applied from the knife. From left at $t=0s$, $t=0.8s$, $t=1.2s$. Here, we have removed the knife from visualization to show the inserted springs clearly.
• Figure 4: Pre-processing of a tetrahedral mesh (visualized by a single tet) prior to being cut along a given cutting surface (red plane). Given the original mesh, (1) intersecting elements are identified, and (2) duplicated such that the intersection geometry is retained in each part of the cut elements. Virtual nodes are inserted where the edges are intersecting the cutting surface. (3) The original and duplicated elements are connected by springs (green) between the virtual nodes.
• Figure 5: 2D slice of the signed distance field (SDF) for the knife shape (not true to scale), where the color indicates distance $d$. The knife's boundary at $d=0$ is indicated by a solid black contour line. Left: at distances greater than zero the SDF becomes more rounded. Right: distance $d$ along the edge $(p_1,p_2)$ with barycentric edge coordinate $u$ varying between $0$ ($p_1$) and $1$ ($p_2$). The closest point found by Algorithm \ref{['alg:frank-wolfe']} is shown in red. Distances are exemplary and do not represent the actual dimensions used in the simulator (see Appendix \ref{['tab:parameters']}).