Motion planning is a prerequisite capability for the robots for applications in transportation, exploration, and search-and-rescue missions. The overall objective of motion planning is to enable robots to plan the low-level motions needed to accomplish assigned high-level tasks autonomously or under human supervision. According to the lessons we learned from the DARPA Robotics Challenge (DRC), we have high confidence that the motion planner we designed in the DRC is an appropriate and powerful tool to plan the manipulation motions for high degree of freedom
(DOF) robot systems.
This research is aimed at designing and validating a general purpose optimization-based motion planning algorithm for completing practical tasks with humanoid robots. The key features of the planner include flexibility, applicability, reproducibility and reusability for different types of robots.
Through formulating the robot kinematics and dynamics properties, the task requirements and the collision avoidance requirements as the objective and constraint functions in our motion planner, a wide range of optimal, feasible, and collision-free motions can be generated.
Motivated by the human-in-the-loop design methodology, in this thesis, our motion planner provides many levels of human assistance interfaces. A novel concept, called motion template, was designed for operator to assign task level requirements to the motion planner. The motion template is a description to record the relationship between the object configurations and the parameters of the task related constraints. It is irrelevant to the robot properties. Therefore, after extending its functionality, the motion template can be reused and shared between different types of robots.
Taking advantage of the flexibility of the low-level motion planner, the high-level task planning can be improved efficiently. An anytime task planning algorithm for humanoid robots was developed, which can quickly determine the desired standing position and orientation for a given manipulation task by implementing our motion planner as an inverse kinematics solver capable of computing a solution without requiring the robot's feet to remain stationary, can plan the footstep trajectory to the desired standing position, can continuously recalculate the desired standing position with random initial pose during the footsteps executions, and can replan the footstep trajectory if a closer stepping target is found.
Due to the solution for a manipulation task is usually not unique, it provides opportunity to enable the robot to choose different method to solve the same task according to different requirements. Risk is a significant consideration for some task such as high-consequence materials handling and space exploration. A risk-aware compositional autonomy framework for humanoid robots is introduced. The main core of this framework is an evaluation method which can score the risk of the motions generated by the motion planner.The collision risk and the fall risk are took into account in this evaluation method.
The algorithms and approaches introduced in this thesis were demonstrated on the ATLAS and Valkyrie humanoid robots in both simulation environment and real world.
- Professor Taskin Padir (Advisor)
- Professor Dagmar Sternad
- Professor Deniz Erdogmus