How DeepMind Is Reinventing Robotics!

Artificial Intelligence (AI) is a transformative technology. Not only can it enable autonomy and machines that can make intelligent decisions, but it can also even reinvent the technological wheels of various industries. Robotics, being an emergent technology to enable autonomy, AI is a beautiful tool that can help flourish the true capability of robotics technology. And Google's AI partner, DeepMind is reinventing robotics once again.

Today, AI is around us everywhere. From different apps to different devices/gadgets and various services we use, AI mainly integrates with these apps, devices/gadgets, or services. With this, AI provides us a superior experience of use with devices capable of making intelligent decisions and predictions. Moreover, AI is very persistent in modern life, with AI in various voice assistants, recommendation systems in services from e-commerce sites to media consumption platforms, and intelligent solutions to make predictions or autonomous decisions. 

With these services and devices, AI has already become an integral part of our lives. Therefore, it is only natural that industries and companies use AI to boost their company performance on the consumer and product development and innovation front in such a scenario. One of these industries where AI has much potential is the robotics industry. 

Read more: How Deep Learning Boosts Robotic Picking Flexibility

The robotics industry in itself is revolutionary, with capabilities to enable autonomy in industries. However, the endeavors of enterprises and various industries pose a massive challenge for robotics to fulfill them alone. So developers and researchers worldwide are trying to embed AI into robotics technology to usher the robotic industry to a new level. 

With the help of AI, robots will not only be intelligent, but they will also be more capable and efficient. They will be able to form elegant solutions and make intelligent decisions. Moreover, they will be able to control and move a physical body which is very hard to program and build from the ground up. Furthermore, with the decision-making and prediction prowess of the system with convergence of robotics and AI, revolutionary and even unseen developments are possible. 

DeepMind is reinventing robotics, and its developers have certainly caught up with this revolutionary possibility. The search giant Google's AI partner, DeepMind, is now working on this problem of convergence of AI with robotics. Raia Hadsell, the head of robotics at DeepMind, said, "I would say those robotics as a field is probably ten years behind where computer vision is." It demonstrates the lack of distinct development in robotics even when tech-like computer vision embedded in robots is already very far ahead. 

The problem lying here is, though, more complex. Alphabet Inc, the parent company of Google and DeepMind, understands this daunting AI incorporation with robotics. More daunting challenges and longstanding problems remain in the Robotics-AI paradigm alongside challenges of gathering adequate and proper data for various AI algorithms to train and test them.

For instance, problems like training an AI system to learn new tasks without forgetting the old one? How to prepare an AI to apply the skills it knows for a new task? These problems remain primarily unsolved, but DeepMind is reinventing robotics to tackle the issues.

DeepMind is reinventing robotics

DeepMind is mainly successful with its previous endeavors with AlphaGO, WaveRNN, AlphaStar, and AlphaFold. However, with various breakthroughs and revolutionary developments, DeepMind is now turning towards these more complex problems with AI and Robotics.

However, a more fundamental problem remains in robotics. With their AlphaGO AI, DeepMind is reinventing robotics and successfully trained it through the data from hundreds of thousands of games of Go among humans. Apart from this, additional data with millions of games of AlphaGO AI playing with itself was also in use for its training.

Read more: Top 10 Companies To Dominate The Future Of Industrial Robotics Market

However, to train a robot, such an abundance of data is not available. Hadsell remarks that this is a huge problem and notes that for AI like AlphaGO, AI can simulate thousands of games in a few minutes with parallel jobs in numerous CPUs. However, for training a robot, for instance, if picking up a cup takes 3 seconds to perform,  it will take a whole minute to just train 20 cases of this action. 

Pair this problem with other problems like the use of bipedal robots to accomplish the same task. You will be dealing with a whole lot more than just picking up the cup. This problem is enormous, even unsolvable, in the physical world. However, OpenAI, an AI research and development company in San Francisco, has found a way out with robotic simulations. 

Since physically training a robot is rigid, slow, and expensive, OpenAI solves this problem using simulation technology. For example, the researchers at OpenAI built a 3D simulation environment to train a robot hand to solve a Rubik's cube. This strategy to train robots in a simulation environment proved fruitful when they installed this AI in a real-world robot hand, and it worked. 

Despite the success of OpenAI, Hudsell notes that the simulations are too perfect. She goes on to explain, "Imagine two robot hands in simulation, trying to put a cellphone together." The robot might eventually succeed with millions of training iterations but with other "hacks" of the perfect simulation environment.

"They might eventually discover that by throwing all the pieces up in the air with exactly the right amount of force. With exactly the right amount of spin, that they can build the cellphone in a few seconds," Hudshell says. The cellphone pieces will fall precisely where the robot wants them, eventually building a phone with this method. It might work in a perfect simulation environment, but this will never work in a complex and messy reality. Hence, the technology still has its limitations.

For now, however, you can settle with random noise and imperfections in the simulations. However, Hudsell explains that "You can add noise and randomness artificially. But no contemporary simulation is good enough to recreate even a small slice of reality truly."

Furthermore, another more profound problem with AI remains. Hadsell says that catastrophic forgetting, an AI problem, is what interests him the most. It is not only a problem in robotics but a complexity in the whole AI paradigm. Simply put, catastrophic forgetting is when an AI learns to perfect some task. It tends to forget it when you train the same AI to perform another task. For instance, an AI that learns to walk perfectly fails when training to pick a cup. 

This problem is a major persistent problem in the Robot-AI paradigm. The whole AI paradigm suffers from this complexity. For instance, you train an AI to distinguish a dog and a cat through computer vision using a picture. However, when you use this same AI to prepare it for classification between a bus and car, all its previous training becomes useless. So now it will train and adjust its "learning" to differentiate between a bus and a car. When it becomes adept in doing so, it may even gain great accuracy. However, at this point, it will lose its previous ability to distinguish between a dog and a cat. Hence, effectively "forgetting" is training.

To work around this problem, Hadsell prefers an approach of elastic weight consolidation. In this approach, you task the AI to assess some essential nodes or weights (in a neural network). Or "learnings" and freeze this "knowledge" to make it interchangeable even if it is training for some other task. For instance, after training an AI to its maximum accuracy for distinguishing between cats, dogs, and you, task the AI to freeze its most important "learnings" or weights that it uses to determine these animals. Hadsell notes that you can even freeze a small number of consequences, say only 5%, and then train the AI for another classification task. This time says for classification of car and a dog.

With this, the AI can effectively learn to perform multiple tasks. Although it may not be perfect, it will still do remarkably better than completely "forgetting," as in the previous case.

However, this also presents another problem: as the AI learns multiple tasks, more and more of its neurons will freeze. As a result, it would create less and less flexibility for the AI to learn something new. Nevertheless, Hudsell this problem is also mitigable by a technique of "progress and compress."

DeepMind is reinventing robotics

After learning new tasks, a neural network AI can freeze its neural network and store it in memory/storage to get ready to learn new jobs in a completely new neural network. Thus, it will enable an AI to utilize knowledge from previous tasks to understand and solve new tasks but will not use knowledge from new functions in its primary operations. 

However, another fundamental problem remains. Suppose you want a robot that can perform multiple tasks and works. In that case, you will have to train the AI inside the robot in each of these tasks separately in a broad range of scenarios, conditions, and environments. However, a general intelligence AI robot that can perform multiple tasks and continuously learn new things is complex and challenging. DeepMind is reinventing robotics and now working continuously to solve these AI-Robot problems. Like DeepMind, FS Studio is also hard at work with its collective experience and knowledge over decades. FS Studio is also improving its services like Robotic Simulation Services, Offline Programming, and Digital Twins for reinventing the paradigm of robotic research and development with AI at its center.

You may also like:

Leave a Reply

Your email address will not be published.