Is a Cambrian Explosion Coming for Robotics? | Hacker News Reader

Is a Cambrian Explosion Coming for Robotics?

pubs.aeaweb.org

57 pointsghosh19 comments

Sorry for the off topic but most non-biologists fail to precisely understand what Cambrian explosion means.

It was a time when suddenly an huge variety of species appeared and what was new among them all was the trait of a body plan or body structure with diferent cells forming organs for specific functions. The explanation most accepted by biologists is the appearance of HOX genes. These are genes that switch on/off the expression of other genes and that is what causes cells in the same organism with the same genetic code to be very different in morphology.

Curious if there's any analogue out there in the current tech world? Printers that can print or assemble themselves? A/B testing coupled to genetic algorithms that can switch off functionality for certain user groups if it's not useful?

Curious that the author suggests an explosive rate of improvement in robot performance but then fails to mention any positive examples of robot performance. Instead the entire article focuses on positive rates of improvement in many constituent technologies needed when building robots.

Judging from the inability of every robot in the most recent DARPA challenge to even get back up after a fall, I'd say today's robots have a long way to go before a Cambrian level of explosive growth is shortly forthcoming.

I too am surprised at the lack of examples. There are numerous.

Robotics used in the automobile industry are extraordinary examples of robotic performance, and have been for decades. They only keep getting better.

Kiva Systems has produced extremely productive robots. Over time (eg 20 years) they're guaranteed to replace a hundred thousand plus human jobs. The concept that Kiva represents of warehouse and stock automation, will replace a lot more jobs than that over that time duration.

Intuitive Surgical's sytems are performance robotics in action. Robotic surgery will get better non-stop for decades to come, eventually being vastly superior as a form of surgery vs traditional non-assisted means.

UAVs are human controlled robotics. They've become increasingly critical platforms to all major militaries world-wide. That will only grow in importance and complexity by the year.

Consumer drones are exploding in popularity, and are another example of human controlled robotics. There will be millions of these in just the US in ten years. The total economic impact will measure in the hundreds of billions of dollars in regards to what drones will affect and improve.

iRobot's home robots (eg the Roomba), and comparable competitor products, have shown dramatic improvements in performance since the day the company was founded.

Every robot but one. The CMU robot got up after a fall.

Edit on why no examples: This isn't a review paper for a popular audience. It's a paper on economic perspectives and he says just what he needs to about where he sees the opportunities. The reader can easily find examples elsewhere.

More optimistic: the first auto-driving-car DARPA challenge was a complete disaster. Just a few years later, they rocked. Why?

Because the 1st test was just the hardware guys putting together the bones, sensors and power plant. Once the software guys had that, then impressive behavior changes started happening.

I think the next time, all the DARPA-challenge bots will be sprinting around, doing gymnastics and talking to you.

The "improvements" and "big ideas" the author listed are real and exciting developments. I think about robotics a lot (in academia) but given these developments, a Cambrian explosion (i interpret as robot succeeding in general simple tasks that human can do) is still way beyond my imagination. There are many technical hurdles such as perception, planning, learning and reasoning. The key challenge now seems to be finding scenarios where robots can be successful with current technology.

On a separate note, the robobrain project http://robobrain.me/ seems to fit the running themes in the paper.

Making better hardware and software for controlling hardware is something that is very difficult (hence most the machines having a hard time with the most basic things for a human - walking, etc.), and because of this, most regard robots as being in their infancy, with many more generations to go before we have a robot that can truly function like a human (in the physical sense).

What's alarming about this is the fact that all of these problems are problems that are solved by intelligence and brute force, and both of these functions can be performed in the virtual world, but a lot of the fears we have about "robocalypse" scenarios are rooted in the physical aspects of robots (e.g., autonomous drones with guns, etc.). Given a computer system with intelligence much greater than a human (not necessarily passing the Turing test or being intelligent by certain human metrics, but being very effective at problem solving and engineering) could result in most of the current problems that companies like Boston Dynamics, et al have been working on for years being solved and tested in simulations in mere minutes. Such a machine can't just be invented, but through evolution (genetic algorithms, etc.) on the copious cloud computing environments, such a machine will eventually happen, unless something changes (e.g., laws to attempt to prevent it, disaster of some sort that sets us back financially and computer-wise a few decades, etc.).

This means, the thing we should really fear is machine intelligence, since the physical world aspect of its existence is really secondary. A machine that can design better solutions to physical world problems (like moving around, negotiating obstacles, etc.) than the hundreds of teams working on the same thing is what I fear.

Before any of this can happen key technologies need to experience massive evolutionary improvements:

    Computers:  
        10x computing power
        1/10 the volume (think: fits in the robot's head)
        1/10 the power consumption of a desktop computer
        10x reduction in failure modes (can remain operational with some damage)
    
    Power:
        2x to 10x the energy storage
        1/2 to 1/10 the volume
        Less weight
        Electric vehicle market will help here!

    Actuators:
        10x improvement in power-to-weight ratio
        10x to 100x reduction in cost
        10x to 100x reduction in weight
        10x reduction in fragility (think: still works after impact from a fall)
        Moldability (think of the shape of every muscle in your body)

There's more, of course, starting with software, languages, tools, etc.

Lots of numbers there. Compute power: I've heard for years that we're always just 10X away from useful AI. Any evidence its finally true? Why not cloud compute and have your bot just communicate, which means we have enough compute power in an Atom right now?

For all we know, we've had the compute power for years, and we just need to find the algorithm.

I doubt it, but we don't have a solid mathematical basis for what intelligence is. It isn't just a matter of building faster computers, we need the theory too.

I disagree on a couple points about actuators. First off brushless electric motors have pretty amazing power to weight ratios, problem is that they have really low torque to weight ratios. We don't need moldability if we take what we learned from electrical engineering and standardize everything.

Cost is a huge issue though, but if we standardize and mass produce them then the problem could be solved. Series elastic actuators have solved some of the fragility problem.

I am talking about the finished actuato rather than just the motor. This includes actuator electromics. Think of all the muscles required to operate a human arm, fromthe shoulder to the fingertip, again, including the electronics to drive them. And then build that into a robot. And then have that robot do gymnastics, MMA, compete in the American Ninja course, jog, trip and fall, in general, deal with, survive and adapt to impact and minor damage.

In the context of actuators, not motors, we have a ways to go. It's not hopeless, of course, but we have to admit that the biological muscle is a really awesome machine.

Series-elastic actuators don't really deal with fragilityy as much as they provide energy storage along one axis. Think of punching a boxing bag or tripping and falling. The actuators we use today are susceptible to damage from relatively minor impact from any direction outise of where they have compliance built in. I can punch you in the biceps and will not break it.

The big problem in robotics is not sensing, electronics, or logic, but mechanics and integration. Just compare the state of the art with our muscles, or the highly optimized integrations of sensors, logic and actuators that exist in the insect world (many of which in isolation we still haven't reproduced) with anything coming out of a 3D printer and it should be obvious we remain far further from the conditions needed for explosive growth than many appreciate.

I disagree. There are plenty of examples of robot hardware that would work great for valuable applications, but the software and sensing isn't there yet.

It's true that most research robots have crappy mechanics and integration. That's because they know that software is the unsolved problem, and hardware is mainly a matter of throwing resources at it. That can wait until the rest of the technology is ready.

Check out a video of IPI's truck unloading robot: https://www.youtube.com/watch?v=Plo7SH9aBgg That hardware more or less existed in the 90s. With the right software, two of those (one throwing, one catching) could unload a truck way faster than humans. But watching it, 90% of the time is sensing and thinking.

Or this humanoid: https://news.ycombinator.com/item?id=9673386 (video sped up 10x). The hardware could indeed be optimized, but it's good enough that it's not the limiting factor. With sufficiently smart software, that could be the robot butler of the future. If you put automotive-scale engineering resources behind it, the hardware would be slick, lightweight, and rugged.

Those robots, compared to biology, are utterly unimpressive. You aren't going to get a robot butler before you have a robot honeybee, but where is the robot honeybee? It's mechanically limited out of existence.

Software etc. are simpler nearer to useful than our hardware is, but it's the hardware parts that are missing to make use of current software.

It's true that robot honeybees will require super-advanced hardware, because the job of a honeybee is stupendously difficult. Long range, weather, precision landings, surviving collisions, etc. That's a long way off.

Robot butlers won't need very advanced hardware, because we've designed homes to be comfortable and benign. Comfortable temperatures, power available, utensils and appliances designed for ease of use. Same with logistics.

So yes, we're a long way from humanoid hardware that could compete in a rodeo, or win a soccer match, or survive in the wild. But butlering is physically not too demanding.

The wing of a plane is incredibly simple compared to the wing of a bird. No muscles, no bones, no feathers - it doesn't even flap. And yet it gets the job done.

Nature has a lot of path-dependence that can be avoided with design. Think of how much cruft accumulates in a software program after 15 years of continuous evolution, and then multiply that by a million. It's not unreasonable to think that whatever hardware we use to emulate various human tasks might be a lot simpler than the wetware that makes up an actual human.

You are correct, actuators are a huge problem and there isn't as much being done to fix it. Right now one of the biggest problems is that we don't have efficient torque dense electric actuators, that is actuators that can produce large torques and don't weigh much. So you end up with robots that are on a per weight basis weaker than we are.

Actuators also don't control what we want them to control, we really need to be able to control actuator force/torque(or even better impedance/stiffness), but most actuators only control position. In fact, the entire reason the Baxter robot is able to work with humans(and not need an expensive cage) is that it has force controlled compliant actuators.

Reply on news.ycombinator.com