Simulation challenges in robotic grasping

Abstract: Grasping and dexterous manipulation is a huge area in the current robotic research field. Traditionally in industrial environments, robots are customized for a certain task and work well with repetitive movements where the entire process is predetermined and deterministic. The possibility for a robot to adapt to its surroundings and manipulate objects with unknown properties is very limited and requires new models and methods. It is advantageous to explore and test new algorithms and models in a simulated environment before building physical systems. This thesis focuses on making the development, design and control of dexterous robot hands easier by exploring the benefits and challenges to use simulation in Algoryx's physics engine AGX Dynamics for experimentation and evaluation of design, motion planning, contact models, geometry, etc. The goal is to simulate complex grasping situations in AGX Dynamics and to build knowledge about contact mechanics to fully capture the dynamics of grasping in simulation. The work consists of validating the simulation library with regards to fundamental physics characteristics involved in grasping by a series of benchmark tests. The validation process aims to verify the models and numerical methods in AGX Dynamics to ensure sim2real transfer. The work has exposed friction as one of the biggest challenges in simulation. Developments of the current friction models that better represent the reality have been implemented and tested in AGX. They both show promising results for further development.