Artificial intelligence or AI is being developed at breakneck speed. While this is not a particularly bad thing, many people are concerned that we may lose control of AI, if it is not regulated properly. There are some organizations that are monitoring the development of AI, such as Elon Musk founded the AI research organization, OpenAI. The company works on several subsets of AI and most of us know of its achievement to best some of the world’s best doters players. Now, the company has taken a step towards bringing human-level dexterity to the bot by training a pair of neural networks that enable Rubik’s Cube to be solved with a single robot-hand.
The robot-arm developed by the company is called Dactyl and how OpenAI managed to train it is quite interesting. The old way to train a neural network is to allow it to practice a task for years at an accelerated pace. When we talk about learning how to beat opponents in a virtual game, this approach is feasible because the software has to spend time learning at every speed that it needs. However, in the case of Dactile, practicing for years with Rubik’s Cube was not an option and, therefore, the company used simulation to train it.
“The system can handle situations never seen during training, such as being hurt by a stuffed giraffe. This suggests that reinforcement learning is not just a tool for virtual tasks, but requires unprecedented dexterity Can solve the problems of the physical-world, “OPENAI says. Dactyl’s dexterity and sharpness is nowhere near humans when solving a cube Yes, it is a great way to demonstrate how simulation can be implemented and lay a foundation for general purpose robots.
According to OpenAI, the biggest challenge he faced while training the network was to create simulation environments that are diverse and unique. The simulations needed to capture real-world physics, including features such as elasticity, mobility, and friction, as well as a model for complex objects such as Rubik’s cubes or robot hands. The company has developed automatic domain randomization (ADR) technology, which is said to generate progressively more difficult environments in simulations.