Welcome to the Bulletin by Remix Robotics, where we share a summary of the week's need-to-know robotics and automation news.

In today's Bulletin -

• More domestic robots

• More Dall-E-2 and why roboticists need to use it as inspiration

• Robot surgeons - Good or Bad?

• Tesla’s Gigapress - Bad…

• Somebody is watching you

• Design vs evolution - why robots cracked art before grasping

Snippets

DOMEstic Bliss - Last week, we saw BEHAVIOR: a virtual environment for testing robots on household tasks. This week, we bring you the equally-questionably-acronymed DOME (Demonstrate Once, iMitate immEdiately). DOME is a new control methodology that has allowed a robot to reach 100% success rate in 7 household tasks after only a single demonstration. You could say that this is one Dome demo that dominates the domestic domain.

Eric Jang on Generative Modelling - The former Google roboticist and current vice president of AI at Halodi Robotics, runs through how researchers can seek inspiration from the rapid progress of generative modelling. Quite technical, and pairs nicely with this week’s Deep Dive. As a bonus, he’s also used Dall-E-2 to generate bespoke stock images (oxymoron?), which exactly match his arguments. Say goodbye to clipart, stock photos and a lot more - Dall-E-2 may be one of the most disruptive advances of the decade.

Determining whether technology lives or dies - A new theory suggests that the successes of an innovative technology may be driven by its compatibility with available software and hardware, rather than its superiority over other technologies. The author calls this a “hardware lottery” and suggests that AI researchers should ignore hardware at their peril - the future’s dominant design will be the one most compatible with current CPU and GPU technology.

Bedside Spanner? - Surgical robots are more accurate and precise than humans, but is their development to the detriment of human surgeons? With each robot designed to perform several different tasks, surgeons aren't learning skills that were required of them in the past. Is this a sign of progress, or are we being stitched up?

Casting Doubt - This is the 3rd bulletin in a row we’re mentioning Tesla’s Gigapress. Unfortunately, 3rd time has not been the charm. A huge proportion (60%!) of rear frame metal castings made at Giga Grunheide are being rejected due to quality issues. The factory has become a “money furnace” as Tesla has halted production at its German plant to resolve the issues.

Feel like you’re being watched? - The Enabot Ebo Air smart robot was found to be vulnerable to being hacked, meaning it could be used as a mobile surveillance device. Whilst there’s no evidence this happened, it’s another wake up call. Are we happy having roving video cameras and microphones walking around our homes? On the other hand, they are very cute…

Unicorn extinction - Fourteen tech companies became unicorns in July 2022, the lowest count since August 2020 with nine companies.

The Big Idea

How Robots Became Artists Before Holding a Brush

How is it possible that robots can generate beautiful art but still can’t pick up a random object? Last week, we discussed the control strategies of flexible robotics and the challenges entailed with grasping. In parallel, it's being announced that AI’s are designing ad campaigns. So, why has it taken physical motion so long to ketch-up?

5 years ago pundits worried that truck drivers, factory workers and restaurant staff were at risk of losing their jobs to automation. No one predicted that artists and designers might be at risk. It seems counter-intuitive that these “high level” activities are the first to be mastered by a computer, even Bezos is surprised -

“I think if you went back in time 30 or 40 years and asked roboticists and computer scientists, people working on machine learning at the time which problem would be harder to solve: machine vision, natural language understanding or grasping - I think most people would have predicted that we would solve grasping first” - Jeff Bezos, re:Mars 2019

Turns out Bezos is wrong. Scientists have been predicting this since the dawn of AI.

AI’s first trip around the hype cycle

In the swinging 60s, the early pioneers of AI were bullish - computers were being used to solve humanity’s hardest challenges - geometry, algebra, chess, etc. In their minds, a general AI could only be a decade away. With the hard problems out of the way, it shouldn't be long before AIs would master easy tasks like vision and basic common sense. Of course, general AI is still a decade away and the easy tasks remain unsolved.

The gap between expectation and reality caused a huge amount of frustration and lead to the “AI winter” of the 70s. According to Rodney Brooks, one of those early pioneers, the issue was that AI researchers had defined intelligence as “the things highly educated male scientists found challenging” whereas -

"The things that children of four or five years could do effortlessly, such as visually distinguishing between a coffee cup and a chair, or walking around on two legs, or finding their way from their bedroom to the living room were not thought of as activities requiring intelligence.”

Based on these experiences, one of Brook’s colleagues coined the first paradox of artificial intelligence -

It is easy to train computers to do things that humans find hard, like mathematics and logic, and it is hard to train them to do things humans find easy, like walking and image recognition. - Moravec's Paradox

Evolution trumps design

Hans Moravec reasoned that this paradox wasn't arbitrary but was driven by the difference in evolution and design.

Our abilities have evolved over thousands of years through a process of natural selection. This process is optimised to produce traits that increase the chances of survival while minimising energy use. In our slow progress from single-celled organism to human, evolution favoured the skills at the top of the survival hierarchy: the ability to sense our environment, manipulate objects and move around. Two million years ago, early hominids needed to run away from sabre-tooth tigers and to find food — not solve chess puzzles. As a result, the ability to reason abstractly didn't evolve until around 100,000 years ago.

Things we find challenging like chess and calculus generally involve abstract reasoning. Often they feel hard because we’re consciously trying to solve them but actually involve quite simple, linear thinking. Evolution hasn't had enough time to optimise these tasks and bring them deeper into our subconscious. Things we find easy like walking, recognising faces, and manipulating tools have been developed over millions of years of evolution. Our brains and bodies have been tightly optimised to make these important tasks feel easy and unconscious.

The older a capability is, the more time natural selection has had to improve its execution and the harder it is for us to reverse engineer. With abstract reasoning, we’re competing with 100,000s of years of optimisation; with grasping or sight, we’re competing with millions. Using sight as an example:

“In addition to the light detectors, the retina contains edge and motion detection circuits, grouped in a small area two centimetres wide and ten centimetres wide that simultaneously reports over one million regions of images about ten times per second through the optic nerve…. In robotic vision, similar detections require the execution of several hundred computer instructions, which causes the 10 million detections per second of the retina to involve more than 1,000 million instructions per second.” - Hans Moravec, 2000

Why Moravec may be wrong

James Lucassen claims that Moravec's paradox is actually misleading and results from the “Availability Fallacy” where we focus exclusively on what we can recall. This bias means that we tend to focus on memorable data rather than looking objectively at the whole picture.

He argues that humans and computers don't have opposing strengths and weaknesses, they actually have uncorrelated strengths and weaknesses.

Lucassen states that -

Things like single-digit arithmetic, repetitive labour, and drawing simple shapes are easy for both humans and computers. Things like protein folding, the travelling salesman problem, and geopolitical forecasting are difficult for both humans and computers. But these examples aren’t particularly interesting, because they feel obvious.

There may be some truth in this, but his selection of examples seems a bit off. Although no algorithm exists that can solve all travelling salesman problems, in 2006 Bell Labs solved an 85,900-city salesman problem, plus Deepmind recently used AI to solve the folding problem for all proteins…

Another theory is that AI’s success is down to a mixture of luck and incentives. Borrowing from the “Hardware Lottery” theory discussed in this week's Snippet, a “Data Lottery” could be responsible for success in AI. Deep Learning algorithms are dependent on large volumes of data. We previously discussed how the recent breakthroughs in a GPT3 and Dall-E 2 were due to the implementation of Large Language models, which use thousands of terabytes of training data. It could be that we only have artistic AI’s because other incentives (ad revenue - thanks, Google) existed to drive companies to organise and tag the world's images.

So what?

AI is mastering image generation because we create and share pictures in huge volumes. Images are also relatively easy to define (as pixels, with certain colours, for example), easy to tag and have a few universal formats. Unfortunately, this quality of data is not available for the biggest challenges in robotics. We’ve discussed how simulation can be used as a workaround but to meet the quality of generative modelling we need to match its data. How can we incentivise the creation and capture of data for robotics?

All of this means that most of AI’s wins have been isolated to the world of bits. It's one thing for a robot to generate a beautiful painting on screen, and something completely different for it to paint it on a canvas. Once again, this highlights the challenges of the world of atoms. To crack grasping not only do you need reams of data and brilliant algorithms, but also novel sensors, actuators, controllers and a myriad of other technologies. Next week, we’ll look at grasping in a bit more detail.

Video

Its All Fun and Games

Last week they automated board games, this weeks its jigsaws!

As always Stuff Made Here videos give you a real insight into how tricky hardware problems can be solved. The solutions may not be robust enough for industry but it show’s you the iterative nature of prototyping and the challenges of “integration hell”. It doesn't matter how well things work independently they will never work together first time. Highlights include how to -

• Use a “Core XY” belt set up for your gantry system

• Use tension cables to stiffen a structure

• Calibrate vision systems for distortion

• A multiplex system to simplify the storage of competes

Tweet of the Week