How To Train Your Robot

Robots might be clever, but teaching them hasn’t been easy. . . until now

Whether the advanced robots of today work in the warehouses of Ocado and Amazon or at the shop front of electronics store Best Buy, it’s clear that more and more businesses are seeing the merit of installing intelligent machines. But while these industry giants have the resources and skills to make robots a reality, others are not quite so well equipped. No matter how many amazing things robots can do, there are often teething problems with the technology. The main barriers to adoption include the time and expertise needed to programme bots to complete even simple tasks. Luckily, Californian startup Embodied Intelligence wants to change this by expanding robots’ ability to learn. The most exciting part, according to founder Pieter Abbeel, is that it will all happen in Virtual Reality.

Monkey see, monkey do
If you wanted to teach a robot how to make a cup of tea, you could do so in a variety of ways. One would be through programming, which is all well and good if you’ve got a background in coding and time on your hands. Another method is reinforcement learning, which rewards robots for making successful attempts at specific tasks. Again, this is something that can take a considerable amount of time, and it’s not exactly exciting. A third option is imitation learning, whereby a human operator demonstrates a skill that the robot simply copies.

What Embodied Intelligence wants to do is combine imitation and reinforcement – but they want to do it in VR. The technology has already proven its worth as a human training tool, creating low cost, immersive environments where tasks are carried out without expending physical resources. Embodied Intelligence’s training involves human teleoperators, who perform tasks or skills that are observed and then repeated by a robot. The machine learning algorithms that guide the robot within the virtual world can mimic human movements exactly, creating a more precise repetition. After a process of gradual refinement, the robot can become even better than the original demonstration. Embodied Intelligence aren’t the first company to train robots using VR. Earlier this year, OpenAI began using ‘one shot imitation learning’, in which robots watch and copy virtual demonstrations. Even though it’s been done before, new companies are emerging to develop the technology further and, as a result, redefine robot learning.

How will VR training disrupt robotics?
VR is an important enabler for communication and collaboration within businesses, creating connections in shared spaces. As our world moves into the digital sphere, virtual interactions are becoming increasingly important. But there’s a fundamental difference between training a human and training a robot, even if you use the same medium. Most people have limited concentration and memories but robots can keep learning, adding skills to a constantly expanding arsenal. Perhaps the most immediate disruption to advanced robotics would be the ability for anyone with a VR headset and controllers to ‘teach’ a machine. Although this is positive in terms of democratising programming, giving certain people access to that power could be disastrous. If robots could access an entire world of virtual training, they could presumably watch any demonstration within it – and there’s no guarantee they’ll always be benevolent. Fuelled by the open source movement, the sheer amount of content available to developers is astounding. If it only took one robot to carry out an inventory checks, transport products, clean the workspace and make a mean cup of tea, then the cost of robotics would fall considerably. SMEs would be able to afford the technology, leading to wider adoption and exposure. But, as always, this brings us back to the exhaustive but completely necessary question of unemployment. It’s worrying enough that robots are already better at certain things than we are, but if anyone can teach a robot to do anything then the future looks bleak for all employees with a heartbeat. At the moment, industrial robots are watching virtual demonstrations to assemble electronic parts. In future, could a humanoid watch a demonstration of compassion to become a better carebot?

If Embodied Intelligence and OpenAI are successful in training bots with VR on a mass scale, the adoption of advanced robotics could spike dramatically. From a business perspective, investing in one robot that can complete multiple useful tasks is a no brainer. However, increasing accessibility could have a negative effect on skilled coders and programmers. Today, developers work hard to allow robots to perform menial tasks. Once an intelligent machine can learn to do anything within a virtual environment, then anybody could equip their robot with whatever skills they chose. Before this becomes the case, safeguards need to be put in place to monitor robot skills acquisition. The Swiss Army Robot is coming, and we need to use it wisely.

One day, could a single robot be responsible for multiple tasks within its working environment? Should anybody be able to train a robot? Will the simplification of programming backfire on technological professionals? Share your thoughts and opinions.