Moravec’s Paradox And The Age Of Robotics

It was striking that Rishi Sunak chose to start the penultimate day of his somewhat challenging re-election campaign at an Ocado warehouse staffed by robots. Watching the Prime Minster surrounded by machines zooming about their business focused the mind on the fact that – whatever your thoughts on his broader premiership – Sunak speaks credibly on tech and the need to shape a vision of Britain that embraces the opportunities presented by technological change. It also highlighted what may be the next frontier in the AI revolution: robotics.

In a recent WIRED interview with Jensen Huang, the CEO of NVIDIA was asked which of the many areas that his company was currently researching excited him most; what was going to “change everything”? Huang’s answer was definitive: robotics. “If you could generate text, if you could generate images, can you also generate motion? The answer is probably yes. And then if you can generate motion, you can understand intent and generate a generalized version of articulation. Therefore, humanoid robotics should be right around the corner.”

It's always fun to look at how previous generations imagined the future – both what they go right and what they got wrong. I always reach for Anthony Trollope’s single work of science fiction, The Fixed Period (1882), which imagines mid-20th century life to be a place of mobile phones and euthanasia clinics, but also one in which steam and horse power dominate and there are no airplanes. Perhaps more than anything, though, the thing that visions of the future got most wrong was that they imagined developments in artificial intelligence and robotics moving in lock step, rather than the former evolving far ahead of the latter. Whether it’s the replicants in Blade Runner or the robots in the work of Isaac Asimov, Philip K. Dick and Arthur C. Clarke, there was a sense that technological “minds” and “bodies” would make their leaps forward together.

The fact that they haven’t is down to something called Moravec’s Paradox, a term coined by the professor of robotics and futurologist Hans Moravec in the 1980s. Moravec was one of the first to identify that while a computer can beat the world’s most brilliant chess player, can perform any number of sophisticated tasks of logic and even – increasingly – creativity, the ability of technology to perform even the most basic physical tasks is significantly limited. We’ve all seen the clips of Boston Dynamics’ Atlas robots performing impressive feats of contortion, but it remains the fact that when it comes to walking, recognizing faces, catching a ball or handling a pair of scissors, robots have yet to achieve anything close to human capabilities.

Estimates for the growth of the robotics market are ambitious, notwithstanding the fact that there has been relatively little to justify this optimism so far. And yet – as Huang’s WIRED interview makes clear – it may be that we are about to enter the age of the robot. Crucially, this will be as a result of leaps forward in AI. The problem with the robotics market to date has been one of lead times and capacity. Robots take a long time to put into production, and often by the time the source code has been written and the machines come off the production line, the problems that they were created to solve have moved on. Robots also face the same issues as EVs and virtual reality: energy density. In order to carry a battery large enough to power it, a robot has to be bulky and cumbersome. Even basic sensory-motor skills require massive amounts of processing power, generating heat and requiring cooling systems that further complicate production.

AI and machine learning mean that robots can learn and evolve as they work. AI is also already being used to drive algorithmic efficiency in robots (for instance in the Ocado warehouse, or in Maruti Suzuki’s Manesar and Gurgaon car plants). Artificial intelligence guides robots to conserve energy and follow more efficient paths; deep reinforcement learning means that these efficiency gains are exponential. In a nice piece of recursiveness, the AI even learns to carry out neural network pruning, reducing the power intensity of its own processing. AI is helping overcome Moravec’s Paradox in other, more practical ways. Convolutional neural networks (CNNs) enhance image and video recognition, aiding robotic perception. AI’s ability to process and react to vast amounts of data may be the answer to challenges around recognition and reaction times.

I wrote a few months back about the narrative surrounding AI’s lack of a killer app. Similar questions swirl around robotics. It’s clear that robots are already being used in a number of important industrial applications, although these are so far relatively unsophisticated in the complexity and dexterity of the tasks performed. For instance, in their automotive applications, robots carry out basic welding and painting jobs, but humans take care of anything with any level of intricacy. Robo vacuum cleaners and lawnmowers are a canard. Speaking to Sumant Wahi and his team, who spend a great deal of time looking at this market, it’s clear that the important use cases will be in increasingly complex forms of manufacturing, drug design and other healthcare applications. Areas where the coming together of AI and robotics will drive genuine improvements in the quality of human life.

So maybe the answer to the quest for a killer app for AI is robotics. And the answer to robotics’ killer app conundrum is AI and the way it will drive forward innovation in the space. It may also be the case that robotics provides the final piece in the puzzle when it comes to Artificial General Intelligence (AGI).

A few years ago, I read a superb book that has stayed with me. The Power of Not Thinking, by the anthropologist Simon Roberts, is one of those works that seems to summon the future, to engage in a deep and powerful way with issues whose importance only becomes really clear later, in retrospect. The book is about technology and embodiment, exploring the ways in which computers that engage with our minds only capture a fraction of human experience, which is located in the body. It highlights the way that technology has radically changed the Cartesian view of the body/mind divide and sets out a number of striking observations and challenges for a more embodied approach to the world of artificial intelligence. In Roberts's vision, AGI will only be possible when AI recognizes that we think with our bodies as much as our brains.

It's perhaps here that the greatest opportunity for both AI and robotics lies. The marriage of artificial intelligence and robotics moves technology to a whole new ontological realm, and could be radically transformative. In looking for use cases to justify valuations in AI and AI-adjacent companies, we may have been too narrow in our focus. It's what AI can enable in parallel technologies that will really change the world and its economies. Robotics may the first place that this observation takes flight.