Bimanual robotic manipulation based on potential fields

Dual manipulation is a natural skill for humans but not so easy to achieve for a robot. The presence of two end effectors implies the need to consider the temporal and spatial constraints they generate while moving together. Consequently, synchronization between the arms is required to perform coordinated actions (e.g., lifting a box) and to avoid self-collision between the manipulators. Moreover, the challenges increase in dynamic environments, where the arms must be able to respond quickly to changes in the position of obstacles or target objects. To meet these demands, approaches like optimization-based motion planners and imitation learning can be employed but they have limitations such as high computational costs, or the need to create a large dataset. Sampling-based motion planners can be a viable solution thanks to their speed and low computational costs but, in their basic implementation, the environment is assumed to be static. An alternative approach relies on improved Artificial Potential Fields (APF). They are intuitive, with low computational, and, most importantly, can be used in dynamic environments. However, they do not have the precision to perform manipulation actions, and dynamic goals are not considered. This thesis proposes a system for bimanual robotic manipulation based on a combination of improved Artificial Potential Fields (APF) and the sampling-based motion planner RRTConnect. The basic idea is to use improved APF to bring the end effectors near their target goal while reacting to changes in the surrounding environment. Only then RRTConnect is triggered to perform the manipulation task. In this way, it is possible to take advantage of the strengths of both methods. To improve this system APF have been extended to consider dynamic goals and a self-collision avoidance system has been developed. The conducted experiments demonstrate that the proposed system adeptly responds to changes in the position of obstacles and target objects. Moreover, the self-collision avoidance system enables faster dual manipulation routines compared to sequential arm movements.