Dire warnings that robots are lining up to steal our jobs conjure visions of sci-fi androids whose capabilities surpass ours in every way.

Yet today the state of the art in robotics isn’t likely to inspire fear, but more likely mild concern for the bots struggling to stay upright.

Many physical tasks humans take for granted, such as walking and getting a book off a shelf, are anything but simple for a robot. Underpinning each seemingly straightforward human action is a complex mesh of systems that allow people to interact with unpredictable real-world environments with little conscious thought.

That stark disconnect between the abilities of humans and bots is why a researcher on the team that won an Amazon-sponsored robotics contest is skeptical that robotics pose a near-future threat to our jobs.

Professor Oliver Brock is head of the robotics and biology laboratory at the Technical University of Berlin. It was a team from the lab that won Amazon’s inaugural contest to measure bots’ performance in picking items from warehouse shelves – a task currently carried out by human workers.

Massachusetts Institute of Technology economist Erik Brynjolfsson has written several books about how AI-enabled automation could displace workers “a lot faster and more pervasively” than steam-powered machines did during the Industrial Revolution. As evidence of the gathering power of robotics he cites the advances Google has made towards creating a self-driving car and the ease with which Rethink Robotics robotic humanoid torso Baxter can be trained to perform manual tasks.

Brock also believes that robots will one-day be capable of taking on a broad range of human tasks but disagrees with those who see it happening soon.

“I’m not skeptical about it in principle. I’m skeptical about the timescale. I think it’s inevitable that robots at some point will acquire human-level manipulation skills, but whether it’s 10 years, 100 or 1000 years it’s entirely speculation.

“The argument that within the next five years a significant portion of our labour market will be taken over by robots, honestly when I read that I smile. These people should go to a robotics lab and see what robots really can do and I think they would change their mind.”

Robotic roadblocks

Physical tasks that are easier to automate, such as certain manufacturing roles in the automotive industry, are already performed by robots.

“We’re starting to reach the limits of what we can do with these large-scale automation systems, where we can structure the environment to support the production process.”

Taking a step beyond these quick-wins will be particularly challenging, he said, as it will mean putting robots to work in environments that can’t be carefully controlled, which would require a leap in the sophistication of robotic systems.

“This is not continuous change anymore. This is not making a robot go a little bit faster or introducing a little bit more clever design. This is a step change that we’re not ready to take, of having robots autonomously perform tasks that are complex and physical. I think this will happen, but I don’t know when, and in my perspective we’re very far away from that.”

The extent of the hurdles to be overcome before bots can take on such roles are reflected by how poorly the automated pickers performed when it came to recognising and manipulating items in the Amazon contest, he said.

“The competition has shown that one of the tricky parts is to build integrated robotic solutions that work in semi-controlled environments,” he said.

Brock stressed that these robots were working in an environment that had been set up to make the task easier for them, which was far simpler than an actual Amazon warehouse, and many still failed to score.

“The fact that many of the teams crashed or had zero points, points to the fact there is some inherent difficulty in building these systems.”

Start to introduce a level of complexity even approaching that of an Amazon warehouse and the task would become far more difficult a bot to complete.

“Right now we have one shelf and the shelf wasn’t moved, the number of objects was limited, the placement of the objects was favourable to the robot and they tried to have items mostly towards the front of the bin.”

Each robot had to pick 12 of the 27 objects that teams had been shown in advance, with a maximum of four objects in the bin, the area of the shelf containing the object.

“Imagine if these bins were stuffed with a dozen objects and you have to remove some to pick something up underneath. Or imagine if there’s pencils inside a box and you need to pick out 12 from the box. These are manipulation problems that are much, much harder.”

The winning picker fielded by Brock and the Berlin team used a vacuum cleaner tool with a suction cup to pick up items – a choice that was only possible because of the limited number of items in the contest, rather than the millions of items found in an actual warehouse.

“As the situation ends up becoming more and more realistic the difficulties become much, much bigger and I think it’s a long, long way, there’s many more different steps that we will have to take, explore and solve before we get anywhere near the versatility and robustness of a human in doing these things,” he said, adding that it took about 15 minutes for the Berlin team’s bot to pick out 10 items compared to the hundreds of items per hour that humans can pick out.

“It’s very clear that the performance of these systems is no match for humans,” he said.

Even impressive results in teaching robots to perform tasks using techniques such as deep learning should be viewed as a step forward on a long road, rather than a breakthrough.

“The results have shown that deep learning is an interesting approach that should be tried in many situations. Will it by itself lead to a breakthrough and we will all of a sudden see a jump in the capabilities of robots? I don’t think so.

“I don’t think a single thing will click and all of a sudden everything is possible.”

Further reading