Human Dexterity
The development of robotic dexterity and tactile perception has come a long way. Robots that can manipulate objects with the dexterity and accuracy of human hands are the desired outcome. For those who do not know, dexterity refers to skills at doing something, especially with hands.
A ground-breaking work from the Computer Science and Artificial Intelligence Lab (CSAIL) at MIT is at the forefront of this field of study. The group took on the difficult task of contact-rich manipulation, a field in which robots engage in complicated object interactions. “The main challenge for planning through contact is the hybrid nature of contact dynamics,” the study notes.
AI Researchers teaching robots Human Dexterity
AI employs the technique of reinforcement learning to train a model using incentives and penalties. To make the process of perceiving things simpler for living organisms and repeatable by a simple robot, researchers at this institution developed a form of reinforcement learning technique called “smoothing”.
Furthermore, their approach opens the door for more complex manipulation involving several contact points when paired with sampling-based motion planning. Or to put it another way: handling an object with two hands. In comparison to the hours required by conventional RL techniques, their research have shown the ability to generate complex movements in just minutes.
Also Read: United States Called This Day Of August “The Sickest Day” Of Year; But Who So?
Dual-arm tactile robotic system
In parallel, “Bi-Touch,” a ground-breaking dual-arm tactile robotic system, was unveiled by the University of Bristol in the UK. According to the research report, “We propose a suite of bimanual manipulation tasks tailored towards tactile feedback: bi-pushing, bi-reorienting, and bi-gathering.” Through sim-to-real deep reinforcement learning, this system may become proficient at complex manipulation tasks including cooperatively pushing and deftly spinning items.
Researchers from Stanford University on the West Coast are utilizing human video demonstrations to teach machines difficult skills. Their approach avoids the requirement for pricey image translations between the human and robot domains by using masked eye-in-hand camera video.
