The robot hand recently trained by OpenAI team can! Solving Rubik’s Cube with one hand is quite a difficult task for most humans. One who wants to do that first has to learn specific moves, algorithms, and how to manipulate a cube with one hand.
OpenAI is a non-profit organisation co-founded by Elon Musk, which investigates artificial intelligence regarding various aspects of life. They decided to construct a robotic hand which would be able to do complex activities and adjust to numerous situations.
The main goal of the projects was making the robot hand learn (by itself) how to manipulate Rubik’s Cube in various conditions. To explain how difficult the task was, researchers said: “Humans have evolved to be able to operate and manipulate our hands, […] and the robot has to learn all of this from scratch”. OpenAI team trained a pair of neural networks using a reinforcement learning method, which is pretty similar to the way that humans learn. The purpose wasn’t to create a massive algorithm and apply it to a real-world robotic hand but to create an environment in which the device and its software can learn by trial and error.
The researchers met several serious obstacles. For instance, to train neural networks, they had to develop a virtual environment to carry on simulations. To prepare the hand for solving the puzzle in various conditions, they were changing some properties of simulation, such as friction, mass or size of the cube, every time it was successful. Even though the hardware they used wasn’t new, OpenAI developed an innovative approach to general-purpose robots. They let the robotic hand learn by itself in thousands of simulations and then hoped that it could do the same in the real world. It actually worked out, because the robotic hand solves the puzzle 20% of the time when the maximum scramble of the cube is applied and 60% for simpler scrambles. It seems like not a lot, but in reality, it is a great achievement.
Image tweeted by Gary Marcus
Article about the whole project published on openai.com opened a huge debate in the AI community. Although the research has many supporters, it sparked some controversy. For instance, Gary Marcus, a cognitive scientist, criticised OpenAI’s project on twitter. He claimes that the post on the OpenAI’s blog is misleading and he is pretty right. Not only does he criticise low rate of successes, but also that the hand isn’t compatible with other objects. He claims that the first word of the project’s title “Solving Rubik’s cube with robot hand” should be “Manipulating”, because the robot hand did not figure out how to solve the puzzle, but only how to manipulate the object. However, there is no doubt that this project definitely plays a role in the development of artificial intelligence. How cool would it be, if general-purpose robots could not only solve Rubik’s cubes but also chop vegetables or even cook a meal?
references:
https://openai.com/blog/solving-rubiks-cube/
https://en.wikipedia.org/wiki/OpenAI#GPT-2
https://www.theverge.com/2019/10/15/20914575/openai-dactyl-robotic-hand-rubiks-cube-one-handed-solve-dexterity-ai
https://www.designnews.com/electronics-test/rubiks-cube-solving-robot-hand-sparks-debate-ai-community/149764363861759
Indeed, this project plays a role in the development of artificial intelligence. Many different job positions could be replaced by this robot-hand in the future.
Indeed, this is a big step for artificial intelligence. Many different job positions could be replaced by this robot-hand in the future.
Firstly, I do not agree with Mr. Gary Marcus. This is a tremendous accomplishment, and we shouldn’t demote their success. Secondly, it certainly would be cool to see artificial robotic hand do tasks that we are not accustomed to. This is a cool idea and I will be paying more attention to their projects.
Yeah, but tbh when reading the article on Openai.com I felt like they literally are telling a story how their robot hand figured out the way of solving a Rubik’s cube, which is not true, so Marcus’ opinion may be justified. However, OpenAI is putting a great effort into researching AI and implementing it into the everyday life of humans, which I admire a lot. What is another cool thing about them is that they are non-profit, and their mission is to help humanity accept the presence of AI.
A fantastic project, not only because of the complexity of the venture but also for the reason that OpenAI developed new learning methods for learning models. Meta-Learning is the most promising paradigm to advance the state-of-the-art of Deep Learning and Artificial Intelligence.
OpenAI caused a lot of discussion by demonstrating ground-breaking capabilities of a robotic hand trained with Reinforcement Learning. Mainly because of Automatic Domain Randomization (ADR), which is a totally new approach. It will be interesting to see where this new ADR method goes next. OpenAI believes that human-level dexterity is on the path towards building general-purpose robots, so seeing one like that chopping plants seems to be feasible.