Learning to Select Robotic Grasps Using Vision on the Stanford Artificial Intelligence Robot

As the field of artificial intelligence becomes increasingly advanced and integrated, it is time to revisit the half-century old “AIDream,” where intelligent robotic agents were envisioned to interact with the general human population. To this end, the Stanford Artificial Intelligence Robot (STAIR) project aims to introduce robots into home and office environments, where they will facilitate and cooperate with people directly. In order for robots to have any non-trivial use in such environments, they must have the ability to manipulate objects, which is provided through robotic arms. An arm usually has a manipulator “hand” attached at the end to allow finer manipulation and, more importantly, grasping. The ability to grasp is crucial; if we were unable to grasp with our hands, we would find it very difficult to perform essential tasks such as eating, and more complex actions such as cooking and working in an office would definitely be unachievable. A robust and infallible grasping system is therefore necessary for STAIR to achieve its goal. In this paper, a novel approach for robotic grasping will be discussed. By considering information acquired from our 3-D visual sensors, we developed a reliable and efficient grasping system for STAIR that works in unknown and cluttered environments. Background The problem of robotic grasping has existed and has been well studied over the past few decades. The conventional approach use the forces applied by the fingers on the object at their contact points to determine whether a stable grasp can be achieved1. While in theory this fully determines the result of the grasp, this approach is not practical because a complete and precise model of the target object is necessary. If the model was inaccurate, force computations would likely be incorrect. When working in unknown and dynamic real-world environments, STAIR can only acquire a model of the environment through visual perception, which is subject to inaccuracies and incompleteness. In practice, applying force computations directly on these models leads to poor results. The limitations imposed by perception have spurred interest over the past two decades in vision-based grasping systems. In particular, it has been found that perception of 2-D planar objects usually suffers from fewer problems. For such objects, the object Lawson Wong1 Learning to Select Robotic Grasps Using Vision on the Stanford Artificial Intelligence Robot