GPU-accelerated robot trajectory optimization using direct collocation method

Abstract

Space robotics involves using robotic systems to carry out tasks in space, either autonomously or with human cooperation. The tasks can vary from exploration and planetary surface analysis to the capturing and deployment of payloads, as well as maintenance of spacecraft. Among the various types, while the articulated robots excel at precise tasks like repairs, assembly, and sampling, the bipedal robots are designed for mobility, assistance, and navigating complex terrains, such as that on the Moon or Mars. Owing to their versatility, articulated robotic manipulators are quite popular and have been deployed in some previous and ongoing space missions such as the Canadarm series that was used on the Space shuttle, and European Robotic Arm (ERA) currently aboard the International Space Station (ISS). Humanoid robots, on the other hand, are currently actively researched and tested for potential use in future space missions. Notable examples include Robonaut and RoboSimian, both being developed by NASA. Trajectory optimization is an integral part of robotic motion planning and control, which helps to find open-loop solutions, including the required control sequence for performing a desired motion task. The direct collocation method is a powerful non-linear optimization technique that transcribes the continuoustime trajectory optimization problem into a non-linear programming (NLP) problem. Due to its simultaneous nature, it can provide faster results, especially by exploiting the sparsity of the problem and taking advantage of parallel processing capabilities of a graphical processing unit (GPU). The main objectives of this thesis work are - Devise a pipeline using the Hermite-Simpson direct collocation method and NLP solver Interior Point OPTimizer (IPOPT) software to get optimal solutions for a given motion task. - Formulate NLP, including all its constraints for an elastic bipedal robot performing two motion tasksreconfiguration in zero-G and jumping in 1-G. - Enhance computation speed of objective and constraints functions along with their gradients/jacobians using GPU and exploiting the sparsity of the problem. - Make use of our pipeline to improve the feasibility and periodicity of a provided sample solution obtained with reinforcement learning for the jumping task. - Investigate the suitability of the method for online re-planning of tasks.