OpenAI is has a reputation for choosing important problems to solve and showing the results in an impressive way. This time, the American Corporation decided to build and train a mechanical hand to solve the Rubik’s Cube, the world-renowned 3-D combination puzzle invented in 1974 by Hungarian sculptor and professor of architecture Ernő Rubik. Several robots have already achieved this feat, but these were robots built or coded without the robots actually ‘learning’. Guiness World Records recognizes a system that did the job in 0.637 seconds. Independent engineers later managed to reduce this to 0.38 seconds. Later, Universal Robots programmed two industrial-bot arms to collaborate on the task.
OpenAI’s researchers wanted to take things one step further. First, they designed a simulated environment where a virtual hand could manipulate a virtual cube. This facilitated a scenario in which the model was able to spend the equivalent of 13,000 years learning to twist and solve simulated Rubik’s cubes. To achieve the main goal, which was to train and teach the robot specific skills, researchers used Kociemba’s algorithm, which would help determine the appropriate sequence of moves. Secondly, they altered the simulation’s parameters from attempt to attempt, for instance, changing the cube’s size or mass, or changing the friction between fingers and cube. This procedure, known as automatic domain randomization, helps facilitate the transition from the simulation environment to the real world. Once trained, the model controlled a physical hand made by Shadow Robot Company, which was modified to feature LEDs to improve motion capture, rubber pads to improve grip, and more durable components.
And the results? The robotic hand had a 60 percent success rate when the cube needed 15 rotations of fewer to solve the puzzle. That rate dropped to 20 percent when the solutions required 26 rotations or more. But as impressive as it sounds, the company overstated its claim to have ‘taught’ the robot to solve Rubik’s Cube. Kociemba’s algorithm, which they used for this project, is more than 10 years old and doesn’t involve learning. Furthermore, the cube has a Bluetooth system inside and motion sensors that tracked its segments. What this means is that the robot hand has ‘only’ been taught to manipulate the puzzle according to an existing algorithm – it has not learned how to actually solve the Rubik’s Cube.
OpenAI certainly achieved one of its goals – everyone is talking about the organization’s success. However, it’s important for the public to not be mistaken about the capabilities of machine intelligence and to remember that there’s much still to be done in this area before such impressive claims can be made.
The source: The Batch by deeplearning.ai
Marcin Dąbrowski, CIO at ITMAGINATION, claims that the controversy around this case is irrelevant and that we should instead focus on the advances in technology and be excited about potential applications in the future.
Computer simulation and thousands of hours’ worth of training have enabled a robotic hand to reach a near-human level. It’s yet another example, after the Boston Dynamics robots, that shows us that it will soon be possible for machines to work on tasks that today are almost exclusively performed by humans. What’s more, these robots able to adapt to the conditions and environments that surround them. The fact that the robotic hand was using Kociemba’s algorithm – and not AI – to solve the Rubik’s Cube is less important than the huge achievement that it is to make a robotic hand replicate the intricate moves of a human hand. This is a significant achievement and one that should open our eyes even more to the possibility of robotics performing tasks in a variety of different business, hazardous and home environments.
Learn it. Know it. Done.