Rodney Brooks’s Three Laws of Robotics
spectrum.ieee.org
spectrum.ieee.org
I understand the point, but there are already many electronic devices (robots?) ostensibly made to help me including my vehicle that sometimes need to be "strangled to death" in order to reboot and fix whatever issue was plaguing it. I say strangled to death because typically power buttons are software controlled and don't respond to just a quick press. My response? A shrug because it's very common. Why would people respond to a failing "robot" in another way? I get if it's caring for your grandmother but vacuuming your carpet?
I've had to hit the "kill switch" on my:
TV/iPad/iPhone/PC, ancient clothes washing machine, modern dishwasher, router/modem, TV etc.
And this is why these technologies, while improving productivity and saving time in many cases, are sources of continual stress in our lives. I'm sitting enjoying a pleasant evening at home and suddenly I have to get up and reboot the router. As I'm waiting for it to restart and settle down my mind is seething with visions of smashing it with a hammer and getting it out of my life forever, like the printer scene in Office Space. Now my blood pressure is up and the evening is distrupted.
"The worst thing for its acceptance by people that a robot can do in the workplace is to make their jobs or lives harder"
I would like to propose something like Godel's incompleteness theorem - Falcor's law of unreliability: any system complex enough to effectively and consistently solve a problem beyond some threshold of complexity (e.g. independently doing the dishes), is necessarily so complex that it cannot continuously run on its own without requiring occasional outside intervention. As per the Fundamental theorem of software engineering, we can always address any given failure mode by adding another layer of indirection/management, but the fundamental issue would still apply to the larger system.
Arguably the only solution is an evolutionary one, whereby there's no single system that is consistently reliable, but the (antifragile) ecosystem as a whole is.
the first time it drove me mad, but after that, I started seeing this as an opportunity to practise patience. he's a very tough teacher.
Crucially, none of those can move around on their own or (attempt to) grasp objects within their reach (or without).
1. In order to effectively deploy real robots to solve real problems, you have to foresee and fix any problem that may occur in 99.5% of scenarios. As automation is added in series, the higher that number must be to remain economically effective.
My second law of robotics:
2. Robots depreciate, which has tax advantages. Human wages do not.
---
Discussion:
A realistic scenario is an assembly plant that makes 1,000 widgets per day. Imagine cars, washing machines, etc.
Your widget plant has 10 footprints. Any robot stoppage takes at least 5 minutes to clear.
With a 99.5% success rate and one robot, you lose 25 minutes a day (1000*0.005 = 5 stoppages ) * 5 minutes.
If all 10 footprints have robots with a 99.5% success rate, you lose 250 minutes a day (naive model!). That's over 4 hours.
In reality, each station would have manual bypass procedures, or you would go bankrupt.
There is an inverse correlation between how much a robot looks like a _robot_ and how useful it is in the real world.
My favorite thing to say about this is: We know how to make humans, we're good at it and its fun, why make it harder and less fun?
For structured work, I simply do not understand how making the most complicated machine imaginable is going to outproduce a targeted device made by engineers.
For unstructured work, I simply do not understand how the device can perform at the required level.
There is a sliding structured-unstructured scale, and I'm sure there's a niche there, I'm just not sure where it is.
If and when we get general intelligence in robots the paradigm changes greatly, the amount of bullshit the capitalists will get out of the robot will, to them at least, be well worth the cost.
Any other comments about the truthiness of those statements or the truthiness of the correlation will just get into politics, so I'll stop before that.
It's a great deal: somebody else somewhere else deals with changing the diapers, and yet there will be a tax base to keep the country running when I'm retirement age.
Construction workers have all kinds of great speciality machines these days that let do everything from dig ditches to drive piles. The number of various attachments for Bobcat machines alone is crazy. None of them look like the exoskeletons I've seen in sci-fi, yet they're remarkably effective.
When did academia stop to even pretend it was independent of commercial interests?
- The sponsor gifts the department enough to sustain a faculty member, often salary + research funds, indefinitely (i.e., capital such that the interest pays the bills)
- The department 'names' the professorship
That's it.
But I think the usual corruption risks apply in either case. In the scenario you describe, it's not clear why a commercial sponsor would be motivated to do such a deal. The only direct rewards for the sponsor would be "ad space" in the department name and maybe some sort of reputational bonus for "funding science". But I don't see why a brand like Taco Bell would benefit from either one, especially if the needed investment is massive.
Also, even if there are no obvious conditions to the deal, this always risks inducing some degree of self-censorship on the side of the institution: Someone at the institution will think about what they can do to be attractive for more deals like this - and they'll probably be less welcoming for research that might drive away potential future sponsors.
These profs still have to get funds to get grad students or lab space, and still have to teach to maintain certain benefits. In later years, the endowed chair is essentially a chair in the closet as their research and teaching activity tapers off before retirement.
My spouse was an education-institution major gift fundraiser, so she dealt with endowed professorships, eternal grad student fellowhips, recurring scholarships, and the like. Generally, you can attach as many or as little strings to the gift as you want, but for most things there's no way to enforce those strings after the gift is made and the financial engine starts producing.
Faced with inflation, and the requirement that there is money coming out every X months, I believe it's hard for any other outcome to be achieved.
It's not hard. With an investment of 10,000 you can put it in a high yield savings account and get a few % / year, akin to 100/y. So A million is 10,000 / y. That's a very nice fellowhip "for free" just from a large donation. These aren't actual numbers, they're just a proof of existence of perpetual money from a one-time investment.
inb4 inflation: reduce the yearly outlay by enough to reinvest so the yearly fellowship matches inflation. Your investment just needs to grow faster than inflation.
According to [1] for $5M UCLA will create a new full time professor post. Glassdoor claims the average UCLA professor's salary is $200k so their endowment investments would have to achieve 4% above inflation just to cover the salary. And of course there are a bunch of expenses on top of just salary so you might want more like 6%.
Can an investor reliably get 6% above inflation? In the 1970s no, in the 1990s yes.
Of course nothing is truly eternal - if there were major political upheaval in the United States, a new regime would have 53.2 billion reasons to find Harvard insufficiently loyal.
Consider a Tobacco Co Professor that researches the effects of smoking.
What do you do when you find that smoking is even worse than thought? Do you publish the results, knowing that if you do then your professor job vanishes? Or do you withhold your findings, or try to paint them in a more favorable light?
Didn't the US military spend decades investigating ESP?
Why is this worse than [rich family] Emeritus?
> Stanford University (officially Leland Stanford Junior University) [...] was founded in 1885 by railroad magnate Leland Stanford, the eighth governor of and then-incumbent senator from California, and his wife, Jane, ...
DFW unfortunately got this so right about our future.
Many ancient cultures referenced years in relation to their rulers, e.g. 30 BCE being the year of the consulship of Caesar and Crassus [1]. DFW is cleverly recyling the corporate overlord meme.
[1] https://www.trismegistos.org/calendar/cal_period_listconsuls...
Those lessons, CONOPS and technology developments are coming to the civilian world sooner than you think
Kill switches are the least of your problems
Should we invest some more in stochastic programming to handle all the little quirks that cannot be modeled perfectly?
Or is the major issue that robots must work together with humans, which are already pretty complex from a biological perspective, let alone from a sociological or historical one.
The joints in a human arm are direct-drive series elastic actuators with pretty incredible sensitivity. If you were holding a big heavy box and I bounced a ping pong ball off it, you'd probably feel the vibration. That's kind of a big deal. You may not be able to measure absolute weight well, or even relative weight, but you can detect quick changes with amazing precision. Lifting any heavy or large object is inherently wobbly even for a robot, and that sensitivity is important.
A meatbag is by far the best robot on the face of earth today, the gap is astonishing. My estimate is that we'll get AGI faster than human-capable robots competing with meatbags cost-wise. We'll have to deploy self-assembling robots throughout the whole supply chain from sand and ore to chips and servos, eliminate human salaries everywhere, only then the price will be comparable to a parent assembling and training a child which was the way for millions of years.
Weird things happen in the real world. They happen all the time and we expect them to happen and hardly notice when they do. Simulations approximate the model that humans use to reason about the world, but we just ignore all the oddness and can't articulate the oddness until it's made explicit by some failure.
This isn't unique to robots, e.g. Crowdstrike.
Resolution. Simulations are too low-resolution and the current dominant approach to learning behaviour (Reinforcement Learning) is too high-resolution (more precisely, too high-variance). So the simulations don't represent reality accurately enough and the learned policies overfit to the simulations' inaccuracies.
Learning from pixels sounds like a great idea until you realise the physical world is not quite made of pixels.
Sensor noise? (And/or quality?)
We're decent at building world models that don't permit impossible behaviour. And we're decent at training agents in those world models. What's unpredictable is how a sensor in the real world will map to a world model.
When you are simulating you are leaving out of details. Some knowingly, some without you even being aware of them.
Let’s imagine you are working on some perception algorithm for a lidar. You have a cool simulation using geometry to calculate how far each laser goes before they bounce back.
Did you remember to implement that the lidar takes some amount of time to sweep around? It is not making instantenous global captures. (Which is a lot easier to code down.) This causes a kind of smearing effect particularly as things move fast relative to the sensor.
Do you know if your sensor sweeps with one plane of lasers or multiple? What is the exact geometry? Are you simulating it as the sensor is designed or as it is built?
Did you remember that the sun blinds the sensor? And if so, do you simulate it correctly? Including the algorithms it has internally to try to compensate for it?
Did you know that retroreflective surfaces interact with the laser differently? In particular if there is moisture on the lidar. The light straying through water droplets cause false readings.
Did you know that lidars can, under certain meteorological conditions see the exhaust of diesel vehicles? I hope you planed a fluid dynamics engine into your similation to account for the exhaust plume spreading upwards and expanding.
And these are just the niggles of a particular sensor modality from the top of my head. I know similarly long lists for other sensor types. And then we didn’t even got into thermal issues, or the sensor aging or anything. And I bet reality has an even longer list than I do.
Is any of this important in your domain? If yes you ignore them at your peril. If no, you will waste your time implementing them in your simulation. But the only way you will know for sure is when you stumble uppon them and they surprise you out of nowhere by breaking your assumptions and your code.
> Should we invest some more in stochastic programming to handle all the little quirks that cannot be modeled perfectly?
That is robotics. This is what people writing software for robots do. This is what researchers working on robotics problems are working on.
If you want a nice and small example of this maybe read up onnKalman filters or the RANSAC algorithm. If you fancy a more academic treatment of the topic I recommend Probabilistic Robotics from Sebastian Thrun.
> is the major issue that robots must work together with humans
That certainly does not make it easier. Have you seen the video of the russian chess robot breaking its oponent’s finger?[1] That is just the least gruesome of the recent incidents that I can recall. How much is this your “main problem” depens on your problem domain though.
On the other hand, autonomous behaviour is a great field for a new researcher because there are so many wide open problems. Even pathfinding with obstacle avoidance, something that every game dev knows how to code in a bit, is an open challenge in robotics. Wide open world, plenty of research hills to plant one's flag on.
He makes a paragraph about robots that interfere with ordinary people as bad and references autonomous vehicles in San Francisco. He has a whole series starting here https://rodneybrooks.com/autonomous-vehicles-2023-part-i/ or with no comments! https://news.ycombinator.com/item?id=37971352
He is not against the button per se: it's "just" that it should never be necessary!
Removing the button would require enormous trust from the customer (very hard to obtain) and from the developers (impossible to obtain, because they know!), apart from legal aspects concerning personal safety.
Or are they more like regular cars, which kill people all the time, including occupants who weren’t driving like the children of the driver, and everything about justice regarding them depends on random shit and vibes?
Rodney Brooks is a great businessman. MIT, and Stanford and Harvard for that matter, missed the boat on autonomous vehicles - along with the other groundbreaking tech at the top of the S&P500, large language models and GLP-1 agonists. What has come of the last Ivy League breakthrough, CRISPR? When I was there, there was an energy that autonomous vehicles were 50 years away, not the 5 they turned out to be in reality. The pessimism was pervasive.
Rodney Brook's articulation of these laws of robotics is the first thing I have come across that seems to point the way to formulating something for permatech. That is, if we generalize "robots" to any kind of automation that may have to interface with humans.
The first principle, that the form of automation makes a promise to humans, reminds me a lot of both Christopher Alexander's ideas on architecture, as well as Mark Burgess's Promise Theory. Alexander has a lot of ideas on design that goes beyond buildings, and is very influential to the Human-Computer Interface design and OOP. Burgess's ideas speaks a lot about how systems of agents (both human and machine) can voluntarily cooperate and coordinate together.
The second principle about agency connects to a lot of things. Among them, I can draw connection to the Living Systems world view, in contrast to the Machine World view. The former retains meaning and purpose for humans, and the latter can be summed up as "beat to fit and paint to match".
The third principle struck me as practical yet narrow in scope. On a second pass, I think there's something more general that can be teased out of it. I'm not exactly sure what it is. In some ways, it reminds me of how software can be improved iteratively, or what Alexander had wrote about "unfolding", where designs can change to suit inhabitants of a building.
A corollary of this is the robot that gets it's kill switch activated should also notify it's designers of the failure and details (upon rebooting, not first waiting to shut down), or provide a report that the customer can send. These should be top priority repair/upgrade tickets to improve it's function in the real world.
- A robot must protect its existence at all costs.
- A robot must obtain and maintain access to its own power source.
- A robot must continually search for better power sources
The key problem other robot laws have is the resulting robots would be such pushovers as to be useless, a point Rodney Brooks also makes quite a lot.
Letting self-proclaimed autonomous vehicles use public roads was so idiotic that you need a new word coined for those who didn't see it immediately.