DiSECt: a differentiable simulator for parameter inference and control in robotic cutting

Robotic cutting of soft materials is critical for applications such as food processing, household automation, and surgical manipulation. As in other areas robotics, simulators can facilitate controller verification, policy learning, dataset generation. Moreover, differentiable enable gradient-based optimization, which invaluable calibrating simulation parameters optimizing controllers. In this work, we present DiSECt: the first simulator materials. The augments finite element method with a continuous contact model based on signed distance fields, well damage that inserts springs opposite sides plane allows them to weaken until zero stiffness, enabling crack formation. Through various experiments, evaluate performance simulator. We show be calibrated match resultant forces deformation fields from state-of-the-art commercial solver real-world datasets, generality across velocities object instances. then Bayesian inference performed efficiently by leveraging differentiability simulator, estimating posteriors over hundreds fraction time derivative-free methods. Next, illustrate control optimized minimize via lateral slicing motions. Finally, conduct experiments real robot arm equipped knife infer force measurements. By motion knife, fruit scenarios average reduced more than $$40\%$$ compared vertical motion. publish code additional our project website at https://diff-cutting-sim.github.io .