They took our jerbs!

One thing that doesn’t get discussed as often is the suppliers and training infrastructure for the new robots as we move into a post-labor economy.  I think the common wisdom is that individual companies will produce and engineer their own solutions, so we expect McDonalds to roll out their own line of automated ordering kiosks, and Amazon to engineer and train their own fleet of warehouse bots (note the shelves designed to interface with these bots, all the same size?  We’ll come back to that).  But this isn’t really the way it happens, new companies generally spring up to design, train and help businesses customize solutions.  For example, companies like Zivelo make off-the-shelf, customizable physical hardware to do signage, ordering and other tasks.  Likewise, Amazon acquired the Massachusetts based Kiva systems to produce their in-house robotics.  So there is at least one market that will thrive in the years to come – producing generally-applicable, multi-use robotics and then training them to the specific needs of a client. But how will this come to pass?

In years gone by, robots were far more well-defined.  Purpose built automatons usually did exactly one function, and hopefully had some sensing systems to decide when to do it and whether they had been successful each time.  As with the robots used to assemble cars, the early models would perform one very specific task – say, welding two parts together for a specific model of car.  If the welder needed to make a different model, or the parts were slightly different the next year, the line would have to be shut down until new patterns could be input.  The need for more flexibility gave rise to robotic automation processes (much of this work happened in the 90s). In this new approach, multiple cars might come down the assembly line and the robot would need to sense which model was in front of it, then do the correct task at the correct time.  This allowed a lot of efficiency benefits to production, but still was more procedural and specific than the AI we think of today, and the robots were still generally single-task devices (like an arm for putting bolts into holes) that could merely do their job in a variety of patterns.

Even this context-dependent thinking couldn’t stretch “old AI” to new-world requirements. In fact, the world of machinery-production managed to hit a snag with this “single-job, limited-context” robotic picture just a little while ago, as the immense degree of customization on some car lines exceeded the ability of 2000-era robots to adapt. Several car manufacturers produce vehicles sufficiently variable on an individual level (from different hubcaps to specific drive train elements) that the current robots can’t really do the right thing every time – or training them to do so with “if this, then do that” logic would be prohibitive and likely buggy.  To adapt to this limitation, other manufacturers have started to use robots in “helper” positions, managing things that would be hard for a person to do and low risk, like holding a part in place for a human to attach.  This represents a move forward to “multi-job, human-defined context” robots, and one could imagine just a few humans commanding and overseeing a fleet of such things, providing context and correction where needed and likewise giving the feedback needed for the robots to continually optimize their behavior.

So what does all this have to do with business opportunity today?  Well, with all this innovation and development, not a lot of attention has been given to smaller scale automation.  It’s my firm belief that 25 years from now, there will be robotic servers at small restaurants, local pizza delivery will happen via drones, and if you go to the hospital, there will be a semi-autonomous robot that checks your blood pressure and brings you meds.  But, the users of those technologies won’t be able to absorb a whole robotics company to turn towards their needs, and they won’t be able to employ a robotics specialist to write custom algorithms or change their entire business infrastructure to revolve around the robots either.  They’ll need totally general, Lieutenant-Commander-Data (or at least Wall-E) style thinking machines to handle specific but open-ended task lists with complex and changing context.  And here’s the kicker – none of that is quite ready to happen yet. It isn’t just the difficulty of making general-task hardware, though that also presents many engineering challenges.  It’s also the training of the existing algorithms that presents the challenges. The earliest robots operated with only the simplest context – “if there is a part in front of me, carry out my one task.”  This was expanded into “if part, sense one of a few enumerated contexts, and choose which of my few tasks to do.”  And with this latest revolution, we’ve managed “act in such a way so as to accomplish several predefined things, e.g. do not hit humans, do not break anything, and accomplish goals I’ve been given like moving a part into a place or getting coffee into a cup without spilling it”.  This is a huge step forward, but it still doesn’t get us to “figure out what a human would want in this general scenario and do that thing.”  That’s at the very edge of reinforcement learning research, and could be the subject of at least a whole article.

To sum up, the biggest players in their respective industries are about to get a powerful advantage with robotics.  As mentioned before, Amazon can afford to purchase a whole company just to make robots for their specific purposes.  It can afford to alter its logistics chain and make new physical factories and distribution centers to revolve around these new robots, with package storage shelves that work in tandem.  In short, it can produce a top-down solution that only really works for itself and would have to change a lot for any other user.  But if I run, say, a florist shop and I want to have a robot helper that chooses flowers, cuts stems, arranges them and puts them gently in a package for shipment, then I have problems.  It’s unlikely that anyone will have a Boston Robotics-style automaton that comes preloaded with an “arrange flowers” routine, and even if it did I’d still somehow need the robot to know when to get what flowers.  A custom robot-solution would perhaps tie the ordering system to the robot, but again, the florist likely doesn’t have the infrastructure or technical knowledge to do this.  So either someone needs to make a company that specializes in making and training “helper routines” to form the brain of your new GeneralBot3000, or someone needs to make an AI sufficiently general that it can watch a human doing their job and “pick up” what happens when and how contexts work.

In my estimation, advancements will happen in that same order, first GeneralBot will come with a small selection of brains and you’ll order the one(s) you want for your set of tasks, and then eventually a truly “general context” mind will be created that will allow watching and then replicating human activities, with knowledge of the surrounding context.  So we’re gonna start with “find dirty dishes in this building, wash them and put them into a pre-designated configuration” and end with “watch your owner get out painting materials, paint part of the house and then listen as they say the whole house needs to be painted every few years.”  And I believe that creating, training and selling pre-made brains will be a big deal starting around 2020 and thereafter.  But long before that, the Amazons, Toyotas and even McDonald’s of the world will have fully workable systems, made possible by their ability to limit context rather than train for the variety. They’ll be able to operate with a lot less labor overhead, whereas others will need to wait for smarter robots to handle their high-complexity environment. Bleak?  I’m honestly not sure.  But I am very sure that we’ll see a lot of jobs changing in the near future, and displacement will be a real issue, especially with large employers.