This is another of Festo's latest robotic inventions and while it isn't as visually stunning as its flying insect, a multi-fingered gripper that learns how to manipulate an object is still impressive.
The LearningGripper is a demonstration of how AI and robotics are coming together. A four-fingered gripper starts out not knowing how to rotate a ball from an initial position to show its logo face up. A learning algorithm, based on reinforcement learning, is used to slowly improve on random attempts to rotate the ball. If the sequence of movements gets the ball close to its final position, then it is rewarded and has a higher chance of being used again.
If you think the task is an easy one, consider having to write a program that operates the 12 pneumatic bellows that move the fingers even to achieve something simpler such as a rotation. Each finger has three degrees of freedom.
The learning algorithm takes about an hour to learn the task, but once it has the program is available for a second gripper to demonstrate.
The potential of robots to move beyond their role in production lines is clear, once the bottleneck of needing to program them explicitly is removed. The mechanical components used to build robots and manipulators are becoming cheaper and systems that learn manipulations are the obvious next step - even if making them work is still difficult.
Everyone who is anyone needs their own neural network package and GNU, not to be left out of the action, has just released version 0.8.0 of Gneural - what else would they call it? Is this just a gestu [ ... ]
The only real problem with the previous release of Google's TensorFlow was that it would only work on one computer. Training neural networks is computationally intensive so the good news is that the l [ ... ]