Robots are dumb. They have no intelligence until a human gives them some. The most used robotic automation is in the automobile manufacturing industry where robots are used extensively for spot welding and painting. The intelligence needed is very basic. Go to location x,y,z and weld for x fractions of a second. Go to location A and spray. When at location B, stop spraying.
Many assembly operations are much more complicated. For example, take a bundle of cables which are connected to a plug, grasp the plug, rotate the plug so it can fit in a socket which has a notch where the pin on the plug is to fit, and insert the plug. An operation such as this could be programmed into a robot, but the programming and testing could take a very long time and may not be precise because the exact position of the wire bundle can vary. An emerging technology in the world of robots includes the ability for humans to train robots.
The new technique is not programming, it is training. Watch the first video and you will see a human holding the robot arm and guiding to do the desired action. The robot will remember what it learned and over time, robots will be able to learn on their own once they are given a goal.
Some very interesting work on robot training is being done at OpenAI. OpenAI is a non-profit AI research company. Their main focus is discovering and enacting the path to safe artificial general intelligence (AGI). AGI is a deep subject. It has to do with when AIs are at least as smart as humans. OpenAI was formed by Elon Musk and a handful of other really smart and wealthy people who are worried about the dark side of AI. OpenAI aims to make it safe so that humans don’t get wiped out by robots in the future. I will be writing more about this in Robot Attitude later. In the meantime, back to training robots.
OpenAI researchers have trained a human-like robot hand to flip a cube with amazing dexterity. Their robot hand system, called Dactyl, is trained entirely from simulation. In other words, the researchers created a computer model to simulate finding the right position of a cube based on a goal. For example, a goal might be flip the cube in the hand until it displays the letter O on the front of the cube and the letter A to the right as shown in the YouTube image above. The result of the simulation is transferred to the robot hand. The hand was effectively given the result of 100 simulated years of training. The implication of this amazing project shows it will be possible to train robots to solve real-world tasks without physically-accurate modeling of the real world. Watch the YouTube video above. Take note of the goals as they pop up on the lower right and then how the robot hand flips the cube to achieve the goal.