Visibility Maximization Controller for Robotic Manipulation
Project lead
Chief investigator
Contributors
Co-supervisors
External collaborators
(QUT) Jesse Haviland, Ben Burgess-Limerick, Peter Corke
Related links
Research topic Perception and Learning, Modelling and Control
Industry application Assistive Robotics
Closed-loop control methods such as visual serving aim to increase the accuracy of the manipulation task by utilizing continuous sensor feedback, and are also robust to dynamically changing environments. These controllers depend on continuous feedback of the target's position, and therefore, losing visibility of the target is detrimental to the controller performance. For eye-to-hand camera setups, self-occlusions are a common problem, which are occlusions caused by a robot's own body.
Using the redundant degrees of freedom of a robotic manipulator, it is possible to perform the task of controlling the manipulator to a target while simultaneously performing a different task, such as avoiding self-occlusions. For complex robotic systems which include a mobile base or a camera mounted onto an independently controllable end-effector, additional redundant degrees of freedom can be exploited to improve controller performance.
We proposed an optimization-based reactive controller, which aims to maximize the visibility to target(s) by minimizing self-occlusions, while simultaneously reaching a goal configuration. An objective function is constructed for a robot to perform a reaching task that aims to keep the target in sight and avoid self-occlusions, while keeping the robot away from singularities and joint limits. The controller then solves for the joint velocities that minimize this objective function at each time-step using a quadratic program.
The performance of the proposed controller is validated in a range of randomized simulation experiments and real-world experiments which include static and mobile bases, static and controlled camera configurations, scenarios with single or multiple, and static or moving objects. These experiments show that the proposed controller successfully reduces self-occlusion rates while remaining robust to complex environments and not sacrificing significantly on task efficiency.