The Artificial Intelligence Research Organization has reached a new milestone in its quest to build OpenAI general-purpose, self-learning robots. The group’s robotics department, the first humanoid robotic arm developed last year, said Doctle learned to solve the Rubik’s Cube with one hand. OpenAI is looking forward to this feat for both robotic add-ons and its own AI software, which allows Docktail to learn new things using virtual simulations, before overcoming a real, physical challenge.
In a performance video featuring Docktail’s new talent, the robotic arm with clumsy yet precise maneuvers leads to a complete cube solution. It takes several minutes, but Docktail can finally solve the puzzle. Movements appear to be less fluid than humans and are particularly unsatisfactory when compared to the blinding speed and raw efficiency of the human speed cube.
But for OpenID, Docktail’s achievement brings one step closer to the much-sought-after goal for the wider AI and robotics industry: a robot that can learn various real-world tasks without having to be trained for months and years, and without the need to be programmed specifically.
Plenty of robots can solve Rubik’s Cubes very fast. The important difference between what they do there and what we do here is that these robots are built with a lot of purposes, ”says Peter Velander, research scientist and robotics lead at OpenAI. “Obviously you can’t use the same robot or the same approach to do another thing. The robotics team at OpenAI has very different ambitions. We are trying to create a general-purpose robot. We are trying to build something more common within its scope, not just a specific task, but how humans and our human hands can do many things. “
Velander has been referring to a range of robots for the past few years that have pushed the Rubik’s Cube solution beyond the confines of human hands and minds. In 2016, semiconductor manufacturer Infineon specifically developed a robot to solve the Rubik’s Cube at superhuman speed, and the bot was able to do so in under a second. It crushed the human world record of sub-five seconds at the time. Two years later, the machine developed by MIT solved a cube in less than 0.4 seconds. In late 2018, the Japanese YouTube channel called Human Controller developed its own self-resolving Rubik’s Cube using a 3D-printed core connected to programmable servo motors.
Over the years, machines have been able to solve the Rubik’s Cube at Superhuman Speeds
In other words, a robot built for a specific task and programmed to perform that task as efficiently as possible is usually best for humans, and Rubik’s Cube Solution software has long since become popular. So developing a robot to solve the cube, even a humanoid, is not so great, and the docktail works at a slower pace.
But OpenAI’s docktail robot and the software that powers it are vastly different in design and utility than the specialized cube-resolving machine. As Wallander notes, OpenAI’s ongoing robotics work is not only about achieving superior results in narrow tasks but also because you need to develop and program a good robot. It can do without modern artificial intelligence.
Instead, Docktail is developed from the ground up as a self-learning robotic hand that reaches new functions just like a human. It has been trained using software that is currently trying to be fundamental to reflect the millions of years of evolution that help children learn to use our hands naturally. One day, OpenI’s hopes will help mankind develop the humanoid robots we know only from science fiction, robots that can safely operate in society without risk, and perform a variety of tasks in a chaotic environment such as city streets and factory floors.
To learn how to solve a Rubik’s Cube with one hand, OpenIII is not explicitly programmed by Docktail to fix the toy; Free software on the Internet can do it for you. It also chose not to program individual movements to do it by hand, because it wanted to detect those movements themselves. Instead, the robotics team gave the underlying software of the hand the ultimate goal of solving the scrambled cube and using modern AI — specifically the brand of incentive-oriented deep learning, called reinforcement learning — to help it find its own way. This is how OpenID has developed its world-class Dota 2 bot to train AI agents.
But until recently, it was much easier to train an AI agent to do real work — playing a computer game, for example — than to train real-world work. Software training can be accelerated to do anything in the virtual world, so that AI can spend thousands of years of training equivalent to just a few months of real-time, thanks to thousands of high-end CPUs and ultra-powerful GPUs.
It is not possible to do the same level of training with a physical robot doing physical work. This is why OpenAI is trying to guide new methods of robotic training using simulated environments in the real world, and the robotics industry is not just experimenting. That way, the software can be widely practiced at several different speeds at the same time, hoping to retain that knowledge once it starts controlling the real robot.
Due to training limitations and obvious security issues, commercially used robots today do not use AI and are instead programmed with very specific instructions. “In the past, the approach was that you use very specialized algorithms to solve tasks, where you have a perfect model of both the robot and the environment in which you are working,” says Welander. “For a factory robot, you have very accurate models and you know the environment in which you operate. You know exactly how to pick a specific piece of it. “
Current robots are far less versatile than humans. To reproduce a robot requires a large amount of time, effort, and money, to do a specific part of an automobile or a computer component. Demonstrate a poorly trained robot with a simple task that has no level of human ability or visual processing, and it fails miserably. With modern AI techniques, robots can be shaped like humans, so that they can use the same intuitive understanding of the world to do everything from opening doors to egg-laying. At least, that’s the dream.
We are still far from that advanced level, and the AI community’s push towards software — such as self-driving cars, machine translation and image recognition — has not exactly translated to the next generation of robots. Currently, OpenAI is just trying to mimic the complexity of a human body part and make that robotic analog work more natural.
This is why the Dactyl is a 24-joint robotic arm instead of the claw or pincer style robotic grippers you see in factories. And for software that powers the dock to learn how to use all of those joints, OpenIII has been simulated through thousands of years of training before attempting a physical cube solution.
“If you train on a real-world robot, what you’re learning is working on what you really want to implement your algorithm. That way, it’s pretty simple. But today’s algorithms need a lot of data. To train a real-world robot, to do anything complicated, you need many years of experience, ”says Welinder. “Even for humans, it takes years, and humans have millions of years of evolution.
However, in a simulation, Wellander said, training can be accelerated just like other things that are popular as in-game and AI benchmarks. “It takes thousands of years to train the algorithm, but it only takes a few days because we can parallelize the training. Virtual training Iki researchers have struggled in the past. In this case, it is really one of the first to be seen in the progress opened said.
When it was given a real cube, Docktail used its training and resolved it itself, and it did so under different circumstances that were never explicitly trained. The cube is constantly interfering with a glove, its two fingers fixed to each other, and members of OpenIII choking on other objects and bathing them with pieces of paper such as balloons and confetti.
“In all of those moves, we found that the robot was still able to successfully turn the Rubik’s Cube. But in training, it’s not so much, ”said Matthias Plappert, a robot team leader of Weilander’s fellow OpenAI. “When we tried it on a physical robot, we were surprised by the visual acuity.”
This is why OpenII sees Docktail’s newly acquired expertise as equally important for both robotic hardware and AI training. Even the world’s most advanced robots, such as humanoid and dog-like bots, developed by industry leader Boston Dynamics, cannot operate autonomously and require extensive work-specific programming and frequent human intervention to perform basic operations.
OpenID, Docktail is a small, crucial step towards some kind of robot that can one day perform manual labor or household chores and work with humans, not in closed environments, without explicit programming to control their actions.
In that vision for the future, the ability of robots to learn new things and adapt to changing environments is all about the flexibility of AI as a physical machine’s vision. “These methods are really starting to prove that they are solutions to managing all the inherent problems and dilemmas of the physical world in which we live,” says Plappert.
