Festo is a company that makes some strange robots that it has been kicking out for years. Several years back it unveiled a robotic BionicKangaroo that uses a pneumatic system to operate. The company is back with a new device called the BionicSoftHand that is a pneumatic robot hand with embedded AI.
The AI inside the hand is used to learn how to grip and handle objects using a method of reinforcement learning called learning by strengthening. This type of learning means the AI doesn’t imitate a precise action; it is given a goal and has to figure things out for itself. That means it uses trial and error long with negative or positive feedback to optimize its actions until it solves the task successfully.
Training of the BionicSoftHand involved giving it a 12-sided cube with one side pointing upwards. Then the AI is trained in a virtual environment to move the cube so that another color is up when complete. The developers say that the digital simulation speeds training considerably.
Festo uses something called massively parallel learning where acquired knowledge is shared with all virtual hands. This means every mistake was only made one time with successful actions available immediately. This sort of training is a big deal because it means that when one of these soft robotic hands anywhere in the world learns a new task, all hands would know the same task.
The RoboticSoftHand is a pneumatic device and uses a 3D textile fabric covering and proportional piezo valves for precise control. The system has the potential for human and robot collaboration thanks to its flexible, pneumatic design and use of lightweight materials. Those same materials will make construction inexpensive and give the robot the ability to grip objects strongly and with sensitivity.