Can we breach Moravec’s Paradox?
The last 12 months have been a bit crazy in the world of artificial intelligence. The public availability of Large Language Models like ChatGPT, Bard, and others has created massive excitement and anxiety around the upsides of and potential downsides of this new breakthrough technology.
In March 2023, OpenAI published a report which showed how GPT 3.5 and GPT4 managed to ace a bunch of exams designed for humans. Given the pace of change, I wouldn’t be surprised if today’s chart has a large number of exams, and the green bars are higher.
(Being a non-native English speaker, I have to admit I was a bit pissed off when GPT could score reasonably well on the GRE verbal exam (>60%), something which I had struggled with when I gave my GRE exams before coming to the United States.
Image source: OpenAI’s “GPT 4.0” article
It would be easy to imagine that if Large Language Models can clear complicated examples like the law exam, can automation of “low-skilled” tasks like picking strawberries by robots be far behind?
The reality is a bit different. Artificial intelligence powered robots can struggle to stack blocks, or pick strawberries efficiently.
According to Taylor Webb, a psychologist at the University of California, Los Angeles, who studies the different ways people and computers solve abstract problems,
“The stuff that these systems are really bad at tend to be things that involve understanding of the actual world, like basic physics or social interactions—things that are second nature for people.”
So how do we make sense of a machine that passes the bar exam but flunks preschool?
So, how can we explain the challenges encountered by ML/AI and huge amounts of computing power to solve seemingly simple problems?
Enter Moravec’s Paradox.
Moravec’s Paradox for Ag Robotics
Moravec's paradox is the observation by artificial intelligence and robotics researchers that, contrary to traditional assumptions, reasoning requires very little computation, but sensorimotor and perception skills require enormous computational resources. The principle was articulated by Hans Moravec, Rodney Brooks, Marvin Minsky and others in the 1980s.
Moravec wrote in 1988, "it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility”
Artificial Intelligence magazine says,
At first glance, one might assume that replicating human intelligence is a matter of emulating the high-level cognitive functions that we associate with intelligent behavior. However, Moravec’s Paradox posits a captivating twist; while machines excel at complex tasks such as mathematical computations or playing chess, they falter at what humans find elementary due to natural selection, such as basic perception skills and sensory skills. These seemingly simple tasks are a result of millions of years of evolution, and they are deeply ingrained in our neural architecture. For machines, these tasks represent an entirely different kind of complexity that is not easily solved with traditional algorithmic approaches.
If we apply this to agriculture, Moravec’s Paradox says it will be relatively easy for computers to solve problems like make prescriptions, analyze seed performance, but hard to solve problems which require perception and fine motor movements like picking a fruit or removing a weed. (And we already know how difficult it is to solve what I called the easy problems above!!)
But why does AI struggle with the simple? The explanation behind Moravec’s paradox revolves around evolution, understanding, and perception. (You can watch this video which explains the 5 different levels of difficulty for robotics)
The skills that we define as ‘simple’ — those we learn instinctively — are products of years and years of evolution. So, while they may appear simple, it’s only because of thousands of years’ worth of tuning.
Things we consider simple are seeing things, recognizing them, lifting them, and moving them around. The complexity of the simple skills we take for granted is invisible, as we have learnt it through thousands of years.
As Erik Brynjolfsson, Director, Stanford Digital Economy Lab pointed out in his fantastic essay, “The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence”
Humans have evolved over millions of years to be able to comfort a baby, navigate a cluttered forest, or pluck the ripest blueberry from a bush, tasks that are difficult if not impossible for current machines. But machines excel when it comes to seeing X-rays, etching millions of transistors on a fragment of silicon, or scanning billions of webpages to find the most relevant one.
Why is it difficult to do agriculture robotics?
In edition 119, “Agriculture Robotics is difficult AF” I had given examples of strawberry picking, which is considered a low to medium skill task for humans.
Picking strawberries at scale, and picking them consistently requires practice and skill. You have to judge if the strawberry is at the right ripeness for picking, grab the stem, twist it, and put the strawberry in your basket, without bruising or damaging the strawberry. Human beings are good at following the process to find strawberries with the right ripeness, grab the stem, twist it, and put it in a container.
Try to build a robot which can do the same job with human or better efficiency, and it is difficult AF.
The difficulty is due to Moravec’s Paradox, and the challenge to train your machine learning, and AI models to learn thousands of years of human evolution in a few years.
In 2022, The Mixing Bowl did a study of the agriculture robotics landscape.
“For the purposes of this robotic landscape analysis, we focused on machines that use hardware and software to perceive surroundings, analyze data and take real-time action on information related to an agricultural crop-related function without human intervention.”
As can be seen from the chart above, vision-aided robotic pickers and vision-aided spot sprayers are some of the hardest problems to solve within agriculture robotics.
The Western Growers harvest automation report from 2021, reached similar conclusions.
Current adoption of automation in harvest and harvest related activities is low, but significant advancements are expected in the next 3-5 years. Due to Morevac’s Paradox, harvest technologies are not as far advanced as pre-harvesting, and harvest assist activities as can be seen from the chart below.
Primarily, the study finds that the overall advancement of harvest automation in the fresh produce industry is so far limited, mainly due to the technical difficulties in replicating the human hand to harvest delicate crops.
Western Growers summary from their 2021 report succinctly highlighted the limits imposed by the paradox..
If going past Moravec’s Paradox requires unwinding and learning the evolutionary process, which has happened over thousands of years, do we have a realistic chance of solving these problems in the near future?
Can we replay the evolutionary learning process to teach computers and robots what humans have learnt over thousands of years?
Are there examples where we have shown drastic improvements in our learning in a very short amount of time - maybe exponential growth, similar to Moore’s Law?
There are other examples where we (as in humans) have made progress faster than Moore’s Law. A striking example is the cost per human genome, which has reduced from $ 100 million to $ 1,000 (and falling) in just 20 years.
Another example is Training Compute (FLOPs) of milestone Machine Learning Systems over time. According to Wikipedia, floating point operations per second (FLOPS, flops or flop/s) is a measure of computer performance, useful in fields of scientific computations that require floating-point calculations.
Source: Compute trends across three eras of machine learning
The report's summary shows extremely fast improvements in computing power over the last 10-15 years.
“Our findings seem consistent with previous work, though they indicate a more moderate scaling of training compute. In particular, we identify an 18-month doubling time between 1952 and 2010, a 6-month doubling time between 2010 and 2022, and a new trend of large-scale models between late 2015 and 2022, which started 2 to 3 orders of magnitude over the previous trend and displays a 10-month doubling time.”
Over the last few years, we are beginning to see AI go past the Moravec’s Parodox. We are beginning to see AI tools like image classification and facial recognition, typically learnt by a child quite naturally.
Will we be able to compress human evolutionary learning in a few years with sophisticated machine learning and AI tools?
Augmentation is on the road to Automation
Should robotics, AI, and ML focus exclusively on automation and replacement of human tasks as performed today? It is tempting to think automation is the holy grail to solve problems associated with human labor, and efficiency.
“A common fallacy is to assume that all or most productivity-enhancing innovations belong in the first category: automation. However, the second category, augmentation, has been far more important throughout most of the past two centuries. One metric of this is the economic value of an hour of human labor. Its market price as measured by median wages has grown more than ten-fold since 1820. An entrepreneur is willing to pay much more for a worker whose capabilities are amplified by a bulldozer than one who can only work with a shovel, let alone with bare hands.” (From Erik Brynjolfsson’s essay)
We have seen the development and adoption of collaborative robots (or cobots) as a way to augment human capabilities rather than to replace them completely.
“Augmenting humans with technology opens an endless frontier of new abilities and opportunities. The set of tasks that humans and machines can do together is undoubtedly much larger than those humans can do alone.”
Source: The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence
Some examples of augmentation in Ag Robotics are new robotic platforms successfully undertaking labor-saving tasks which are not very difficult. For example, the GUSS autonomous sprayer can work in orchards. The GUSS machine navigates autonomously to adjust spraying based on its ultrasonic sensors.
Other examples include Burro’s smart farming system. The system is described as augmenting human capabilities and working side by side with farm workers.
“We've built a smarter farming system using user-friendly, autonomous robots that work side-by-side with farm workers to make agriculture more productive and sustainable.”
This video from Burro shows an automated mower, which just keeps working as long as there is something to mow, and its batteries can be charged.
In summary, Moravec’s Paradox is a real barrier to progress in Ag robotics, but we have made tremendous strides in going past it. Given that augmentation of human capabilities can create far more efficiency than just pure automation, researchers, and entrepreneurs should constantly look for ways to augment, rather than going purely for automation and replacement of human skills.
I am excited about being at FIRA to see different approaches taken by agriculture robotics, and other technology companies. Are they going for augmentation or automation?
I hope to see you at FIRA 2023.
References
1. GPT-4 Technical Report https://doi.org/10.48550/arXiv.2303.08774
2. Parameter, Compute and Data Trends in Machine Learning by Jaime Sevilla, Pablo Villalobos, Juan Felipe Cerón, Matthew Burtell, Lennart Heim, Amogh B. Nanjajjar, Anson Ho, Tamay Besiroglu and Marius Hobbhahn; 2021.