Teaching Robots. . . with Virtual Reality

As virtual worlds improve, so do their uses

Intelligent robots are gradually moving into our homes and workplaces. These highly capable machines are able to perform a variety of tasks, but before they can carry out an action, they need to be taught how to do it. Previously, robots and AI were programmed to complete jobs by following regimented routines. Now, it’s much easier to train robots via demonstration. Rethink Robotics’ Sawyer model can be trained simply by moving its industrial grade arm through whatever action it needs to take, essentially imitating what a human has shown it. However, Elon Musk’s non profit OpenAI has taken learning by imitation to a completely different level, and is currently teaching robots using virtual reality. But what does this mean for the field of advanced robotics, and how will it affect the development of intelligent machines?

Robots that learn like we do

The manual training software used by Rethink Robotics doesn’t require any programming at all, making it less complicated and far quicker. This is exactly the same for OpenAI’s ‘one-shot imitation learning’, in which robots watch a virtual demonstration and then attempt to replicate it. This imitation technique is how children learn to walk and talk. The new process involves trial and error, which is also an integral part of human learning. When it comes to artificially intelligent technology, developers seem obsessed with anthropomorphism, we just keep making machines more human.

AI research focuses on making algorithms as smart as, if not smarter than, human intelligence, which is perhaps why human learning techniques have influenced robot training. The key difference between regular imitation learning and OpenAI’s training is that it takes place in VR, which means that demonstrations can take place in any given environment. In future, robots could even gain skills by watching games instead of demos. Machines could learn basic driving skills by watching Grand Theft Auto, for example, but whether this would make them safe drivers is another matter entirely. As OpenAI is committed to creating friendly artificial general intelligence (AGI), we can at least hope that robots won’t be hijacking cars and going on joy rides.

How will VR training change the way that robots learn?

Although OpenAI’s bot has only learnt to stack blocks so far, the development of one shot imitation learning represents important progression in advanced robotics and artificial intelligence. In short, it’s an incredibly easy way to expand robot capabilities, allowing humans to instruct bots anywhere in the world to cope with whatever they face. This has clear implications for any industry that requires manual labour, including those where social robots are replacing hospitality and customer services staff. Because training takes place in a virtual world, robots can learn how to deal with any environment imaginable. It would be just as simple to show a bot how to stack supermarket shelves as to teach them how to perform seemingly more complex tasks.

The new teaching method enables bots to do far more, especially considering the development of new, self learning algorithms. The ability of machines to share knowledge could mean that instead of teaching a whole fleet of bots, only one would need to watch the demonstration. However, the growth of shared neural networks is an intimidating prospect. Imagine if robots were able to teach each other but used this knowledge defiantly or maliciously. Programmers need to be able to control these exchanges, or we could end up with machines that are far cleverer than we want them to be.

Ultimately, the ability to teach a robot using VR is a promising method that enables machines to absorb information and apply it in any environment. Training in this way is also far quicker than painstaking programming – particularly if one robot could then teach an entire fleet. In future, virtual training could even be used to train the robots that RoboticsX wants to send to Mars. But if developers want to make AI smarter than humans then it might make more sense to stop encouraging human-like learning and strive for something more. The burning issue is whether or not we really want this to happen, especially in light of new predictions that robots and AI will replace workers in half of all jobs.

Could virtual robot training enhance your business? Which industries will benefit most from one-shot imitation learning? Could robots eventually learn from games? Share your comments below.