Benchmark of the Physics Engine MuJoCo and Learning-based Parameter Optimization for Contact-rich Assembly Tasks

Abstract

Digital simulation tools are used throughout the entire product development cycle, from designing components using finite element methods to simulating their assembly and planning their production. Especially in research, physics simulators and engines such as MuJoCo are increasingly used for rapid data generation in training machine learning models. The quality of the training data is crucial for successful training and the subsequent deployment of the models in real-world applications. The presented work reports on a systematic benchmark of the physics engine MuJoCo. Representative processes in robotic assembly applications are carried out in simulation and on a physical testbed. The results are compared, focusing on the robot arm's pose and the acting contact forces during the assembly processes using Minimal Jump Cost. Subsequently, a deep reinforcement learning agent takes over the simulation's parameterization, and a first comprehensive system for the process-specifc training of simulation parameters is developed and evaluated to increase the data quality and to partially encapsulate the setup effort involved in digital simulation, aiding offline design and validation. The approach results in overall 8% less dissimilar simulations but not in all the individual scenarios.