SliceIt! -- A Dual Simulator Framework for Learning Robot Food Slicing

TL;DR

This paper tackles the challenge of safely teaching robots to slice food in shared environments while adapting to varying material properties. It introduces SliceIt!, a real2sim2real framework that couples a high-fidelity cutting simulator (CutSim, based on DiSECt) with a robotics simulator (RoboSim in Gazebo) to enable safe, data-efficient RL training of compliant knife manipulation. The approach combines a forward dynamics compliant controller with Soft Actor-Critic reinforcement learning to learn a policy that outputs a reference trajectory and FDCC gains, calibrated from a few real cuts and deployed on real UR5e arms. Empirical results show reduced contact forces during real slicing and generalization to unseen foods, albeit with higher computation time due to the realistic simulation, underscoring a practical path toward safer, waste-efficient cooking robots.

Abstract

Cooking robots can enhance the home experience by reducing the burden of daily chores. However, these robots must perform their tasks dexterously and safely in shared human environments, especially when handling dangerous tools such as kitchen knives. This study focuses on enabling a robot to autonomously and safely learn food-cutting tasks. More specifically, our goal is to enable a collaborative robot or industrial robot arm to perform food-slicing tasks by adapting to varying material properties using compliance control. Our approach involves using Reinforcement Learning (RL) to train a robot to compliantly manipulate a knife, by reducing the contact forces exerted by the food items and by the cutting board. However, training the robot in the real world can be inefficient, and dangerous, and result in a lot of food waste. Therefore, we proposed SliceIt!, a framework for safely and efficiently learning robot food-slicing tasks in simulation. Following a real2sim2real approach, our framework consists of collecting a few real food slicing data, calibrating our dual simulation environment (a high-fidelity cutting simulator and a robotic simulator), learning compliant control policies on the calibrated simulation environment, and finally, deploying the policies on the real robot.

Figures (8)

• Figure 1: Overview of proposed learning-based robot cutting framework, comprising four key stages: 1) Data collection on the real robot. 2) Calibration of the cutting simulator, DiSECt. 3) Learning a control policy within a dual simulation environment using Gazebo and DiSECt. 4) Deployment to the real robot.
• Figure 2: ROS-powered proposed system for learning robotic cutting tasks using a dual simulation environment, reinforcement learning, and compliance control.
• Figure 3: Forward Dynamics Compliance Controller scherzinger2017forward
• Figure 4: Framework for learning food slicing using RL and compliance control. The RL agent controls provides the reference trajectory and control parameters for the compliance controller.
• Figure 5: Experimental setup with the real robot.