The Amazon Picking Challenge was a very hard problem for a robot and computer vision systems to solve, it's no wonder they did so badly compared to humans. However, this doesn't mean there's no scope for widening the class of problems that they can tackle. Our recent work in this area suggests that the key to success is to choose the scope carefully – aim for a system that's flexible enough to significantly increase productivity, but not so flexible that it can’t do anything quickly. Another thing we've learnt is that development of these types of system is expensive, so isn't worth doing without a business case in place. The proof of concept stage isn't the hard one – it's relatively straightforward to build something out of off the shelf robotics and cobbled together Matlab that runs a new image processing algorithm... slowly. The hard part is the engineering, not the science – making the whole system run as fast as a human would, or faster, safely and robustly. And that requires big teams of expensive, experienced engineers – it's not going to arise spontaneously from hobbyists or university teams working on a Raspberry Pi.
But – we are starting to see business cases where the rewards for innovation in automation are so great that it's worth investing large sums in. So don't be surprised if robots start popping up in unexpected places in the near future – we're not going to see androids walking the streets in my lifetime, but they do have the potential to transform a whole new set of industrial and commercial processes.