SliceIt!: Simulation-Based Reinforcement Learning for Compliant Robotic Food Slicing

Cristian Camilo Beltran-Hernandez, Nicolas Erbetti, Masashi Hamaya

Keywords: Real2Sim2Real, Reinforcement Learning, Compliance Control

TL;DR: A dual simulator framework to train robotic food slicing skills on simulation then transfer them to the real world

Abstract:

Cooking robots have the potential to greatly enhance the home experience by automating food preparation tasks. However, enabling a robot to safely and dexterously manipulate kitchen tools like knives while handling delicate food items poses significant challenges. This study tackles the problem of training a robot arm to perform robust and compliant slicing motions on food items with varying material properties.
We present SliceIt!, a simulation-based framework for training robust food-slicing skills through reinforcement learning before deployment on the physical robot. Our approach follows a real-to-sim-to-real pipeline: first collecting a small dataset of real food-cutting examples, then calibrating high-fidelity simulations of knife-food cutting interactions and robot motion control. Reinforcement learning agents are trained in this calibrated simulation environment to learn optimal compliance control policies that modulate knife forces. The learned policies are then transferred to the real robot, enabling it to perform intricate food-slicing tasks efficiently and safely by leveraging simulation-based policy training while minimizing real-world training risks, effort, and food waste.

Video Publication Concent: (Recommended) We consent that the video will be uploaded to the workshop's YouTube channel when the submission is accepted.

Submission Number: 2