Numerical Simulations for Physics-based Training of Robots for Manipulation of Flexible Rods*

Robotic manipulation of flexible ropes has wide ranging application from advanced manufacturing to robot-assisted surgery. We report a physics-based scheme to deploy elastic ropes along a prescribed trajectory with a collaborative robot. A numerical simulation tool, based on the Discrete Elastic Rods method, is developed for the modeling of elastic rods including contact and friction. Given the stochastic nature of the friction between the rope and the ground, avoiding friction is a key to repeatability of experiments with the robot. Exploiting the robustness and efficiency of the computational framework, we generate training data in the numerical simulator. This data is used to plan the optimal path for the robotic arm such that friction is minimized during deployment of ropes on a given trajectory. Compared with physics-blind methods that require empirical training by humans for every single rope, our proposed scheme remains valid for any elastic rope regardless of the geometric and material properties. Moreover, vast amount of data can be produced from the simulator in a few hours on a contemporary CPU to train a general neural network with high accuracy.