Robotics grasping

As an intelligent species, humans can perform very complex task easily. However, Humans can not match with robot to do repeated tasks. The robots can be manually programmed to do various tasks. The robots can be manually programmed to do various tasks. However, the working environment around robots is not deterministic. The robots need to be flexible like humans, according to the working environment. This can be achieved using reinforcement learning (RL), just like humans. RL provides humans like psychological and neuroscientific behavior of robots, how to optimize control problem for a given environment. Robots perceive the real-time environment information from sensors. This sensory information is used to train agent. RL agent allows robots to learn tasks by trial and error through interactions with the environment.

During my internship in Fraunhofer, I used state of the art pushing and grasping work done at Princeton. The gripper and objects used during the internship are not shown here, because of company policy. However, this method can be applied to any kind of gripper and objects. The combination of pushing and grasping can help to pick the object from clutter. Consider a scenario where two objects are placed together tightly. In that case traditional methods of grasping fails as when there is no space for the gripper to grasp as they only give the accessible grasp. Experiment setup fro training a model is shown in figure. For the original reference click here.