Comment by iandanforth | Hacker News Reader

Comment by iandanforth | Hacker News Reader

Parent Full thread

It's a big company so it should be no surprise that some sections of Google are very excited about robotics. For example you may have seen https://www.youtube.com/watch?v=iaF43Ze1oeI which is research done by Levine et al at Google. One thing to notice in the video though is that those are all custom arms. Someone was able to get $1M+ to build out that lab and the work is on going.

I think the thing that Google has recognized is that cutting edge robotics research that takes advantage of the work they have already done in deep learning doesn't require the kind of robots that Boston Dynamics (or many of the other companies they bought) build. Cutting edge robotics (from the perspective of learning systems) is back at the level of getting a single arm and a camera to learn basic tasks. In fact, there is a lot of research that can be done in simulation with these new deep methods that doesn't even require hardware.

The software and learning systems they can develop will very likely make their way into super advanced complex hardware like Atlas in the future, but for now it's probably not the best use of their time.

When I was in school in the late 90s it seemed like all of the robotics and AI research were in create algorithms to emulate behavior. Boston Dynamic's work seems to be extended from that. They purpose built a robot with purpose built balancing and walking algorithms. The being pushed over and getting back up demonstration seemed to confirm that to me.

Everything I'm seeing these days with AI research seems to be about developing systems that learn for themselves. Rather than telling the robot/AI exactly how to behave in a given situation, it's about allowing the robot/AI to experience as many situations as possible and learn what the appropriate response should be so that in the future it can independently identify and react accordingly.

I was also in AI school in the 90s and I agree with you to some extent. But when you think about nature, much of the behaviour of animals is not cognitive, but instinctive and physically dependent on their structure. In other words if you can build a robot that knows how to walk because you built it that way, why not do so and build more interesting learning at the decision making level?

A self writing behavior tree that uses behavior to write new behavior based on past behaviors. Sounds human enough to become an addict of some sorts. :-).

Well the problem with that approach is that you have to account for every possible situation that might occur and program them into the robot. When a situation arises that you didn't anticipate then bad things happen. When a learning machine, it can teach itself, adapt and understand.

The former approach works fine for a factory floor robot that is in a controlled environment but doesn't lend itself well to other situations. An example might be a biped robot walking up a mountain and it falls. Most of the examples I've seen, the Boston Dynamics one included, have the robot detect it's falling and put itself into a crash position where it remains until it comes to a rest before attempting to recover.

If your robot is rolling down a mountain it might be destroyed before it comes to a rest. Having a feedback loop, reflex reactions and the ability to access the situation and recover dynamically would be much more useful.

Having a feedback loop, reflex reactions and the ability to access the situation and recover dynamically would be much more useful.

Agreed. But none of that implies learning. So why are you talking about learning in the beginning of your comment?

This happens all too often in AI conversations. Learning gives you a special and powerful kind of flexibility, of course. But not being able to learn doesn't imply it can't cope with an infinite range of situations. A robot that's unable to learn could be programmed with enough flexibility to walk on any surface it could possibly encounter.

>A robot that's unable to learn could be programmed with >enough flexibility to walk on any surface it could possibly >encounter

How exactly?

I meant to say that a learning robot could adapt to a situation and try out possible solutions, measure success and adapt where as a preprogrammed robot would only ever try what it's been infused with. In my analogy of falling down a hill, the learning robot might not be able to stop itself on the first try but hopefully it might adopt a strategy that could allow it to regain control of the situation.

I think it does imply learning. Human babies "learn to walk", by using the feedback from lots of little experiments to improve their ability to navigate uncertain and varying terrain.

Basically, flexibility and the ability to deal with novel situations is close to synonymous with the ability to learn.

Cutting edge robotics (from the perspective of learning systems) is back at the level of getting a single arm and a camera to learn basic tasks.

Which is a nice way to say that robotics really hasn't progressed that much in a sense an outsider would see as progress (where progress on outsider-terms would be Boston-Dynamics style robots that can, say, learn task and repeat them), though I'm sure researchers can point to a lot of progress on their terms.

  Which is a nice way to say that robotics really hasn't 
  progressed that much in a sense an outsider would see as 
  progress

I work in mobile robotics, and I can tell you that there hasn't been a lot of progress.

I wrote a post four years ago about all the various huge showstopping technical challenges facing mobile robots: http://c1qfxugcgy0.tumblr.com/post/31187427192/the-enduring-...

Basically nothing has changed since then.

To have a useful anthropomorphic robot you need better:

  batteries
  actuators
  software
  computers

Only in software and computers are we seeing fast progress. (Big improvements in ConvNet algorithms, and the fast video cards you need to run them) But actuators need to be more powerful and much lighter, and batteries need to store at least ten times as much power. Progress in these fields has been slow, since energy storage is a mature field, and you don't see routine doubling of performance like you do with CPUs.

And, of course, an economically useful anthropomorphic robot has to be dirt cheap, as well.

Batteries have changed a whole lot since the 90's, though. Lithium-ion has a much higher energy density than lead-acid or nickel-cadmium. And without these improvements, many modern robots (roombas, quadcopter drones, etc) would simply not be possible.

Roombas run on NiMH batteries.

The PR2 I discussed in the blog post does indeed have a lithium ion battery pack. It has a 1.3kWh capacity, (188 times bigger than an iPhone 6's battery! Probably part of the reason the robot massed 220 kilos) which gave it a rated... 2 hours of runtime.

Lithium ion is better than earlier battery chemistries, but it's still not very good: https://upload.wikimedia.org/wikipedia/commons/thumb/c/c6/En...

450 lbs, not 480 kgs. The main power draw comes from the 2 desktop computers sitting in its base. Runs for ~4 hrs if you turn one off.

Fixed.

And, of course, an economically useful anthropomorphic robot has to be dirt cheap, as well.

Indeed, people talking about this subject often don't realize that humans are really cheap in many if not most circumstances (Boston Dynamics is working on poison-gas-protection-suit testing robot. Finally figured out a job a person wouldn't do).

And it's an evil equation where once a given task is mastered by robots, it makes humans cheaper in many other tasks - because it increases the competition and because it decreases the cost of maintaining the human.

So we've seen incremental automation and steadily declining living standards. Not a world that screams out the benefits of technology.

Yet another reason basic income would be useful. By leveraging existing capital to establish a more stable floor under that progression.

E.g. "Why would I do that for that much? I don't have any driving physical needs pushing me to poison/injure myself performing a dangerous menial tasks for minimum wage." Which puts a floor on human desire to do basic jobs. Which helps continue to support investment in improving automation/robotics. Which makes the world a better place.

Which actually sounds a lot like a carbon tax and the struggles alternative energy sources have gone through. Call it a self-aware employment tax.

Is it theoretically possible to achieve the kinds of performance needed from batteries and actuators ?

Actuators, yes. At last. The power to weight ratio of motors has improved considerably in recent years. The current record is 5KW/kg [1], for a Siemens motor for aircraft. Tesla's motors are around 3.5KW/kg, although that may be peak, not continuous. Water cooling, as used by Tesla and Schaft, helps a lot. Schaft's innovation was to apply water cooling to small motors.

You can run electric motors far above their continuous rated values for short periods. Also, electric motors specifically designed for brief overloads (high-temperature insulation, temperature sensors) are quite possible. Every automobile starter motor is such a motor. With synchronous brushless motors ("brushless DC" and "variable frequency synchronous" motors are the same thing; motors above a few KW tend to be called the latter) and big power IGBTs, you can have huge torques briefly without much difficulty. If you have the electric power available.

Batteries, maybe. Running time between charges is going to be a problem for a long time to come. There's a huge battery industry trying to get energy density up, with modest success. For many applications, trailing a power cord most of the time is an option. Especially if the robot can plug itself in, which the Hopkins Beast was doing in the 1960s.

[1] http://www.gizmag.com/siemens-world-record-electric-motor-ai...

The power density of electric motors is pretty amazing. Take for instance quadcopters, a couple of brushless motors can lift their own weight, batteries, and still have plenty of thrust left over to accelerate[0]. Yes those power densities are probably for continuous power.

The problem is not power density, but torque density. Brushless motors spin really, really fast with low torque, which is the exact opposite of what we need for robots.

[0]https://www.youtube.com/watch?v=8p5uDf9i_Yc

Electric motors have maximum torque at zero speed. Maximum power is at half of no-load speed.[1] You can design motors for higher torques at lower speeds; it's a standard design parameter.

[1] http://lancet.mit.edu/motors/motors3.html#tscurve

I think the argument here is that this maximum torque is just not enough.

The maximum torque an electric motor can produce is proportional to the magnetic field in the windings which is proportional to number of turns and current. Increasing the number of turns means more mass, increasing the current means more heat. More heat is particularly insidious, because as temperature goes up resistance increases, which makes more heat, which increases temperature, and so forth and so on.

SCHAFT found a solution to the more current problem with their ultracapacitor driven water cooled motors. Except one cannot drive said motors continuously and alone they still don't have that much torque, so if one wants more torque more windings must be added.

You can design motors for higher torques at lower speeds but the torque density suffers. Luckily we have compact high reduction gearing to transform high speed low torque to low speed high torque.

You've made some good points in this thread about actuation. Can you point me to more info on SCHAFT's actuation strategy? I've been trying to turn up info, but haven't found anything great.

>> For many applications, trailing a power cord most of the time is an option.

If the charging time was close to zero, would this problem be solved, for many applications ?

Or the other alternative, phinergy's aluminum air battery, which has 2000 wh/kg, but cannot be recharged, just replaced and "recycled", but probably in a cost effective way ?

Primary batteries have higher energy densities than rechargeable batteries. The military uses high-energy-density primary battery technologies for one-shot items like torpedoes. This might be worthwhile if you were building robots to help decontaminate Fukushima. Commercial applications, no.

Is the hardware really that much of a problem ? It seems like the Boston Dynamics robots are already very far along on the mechanical aspects. I've also seen videos of industrial robots which appear to work with greater speed and agility than humans. My impression is that solving the control problem robustly and rapidly is a bigger obstacle than mechanical limitations.

I do see the power source as being an issue. Many of the most impressive robots are tethered, but even a tethered humanoid robot could be extremely useful.

BigDog uses hydraulic actuators, and is powered by a two stroke gasoline engine. Not so good indoors.

Stationary industrial robots, the only real success story of robotics, have great speed and power, at the cost of incredible weight and power consumption.

Consider the Motoman EPX2050.

http://www.motoman.com/datasheets/EPX2050.pdf

15Kg payload, pretty okay, (Try holding a 15Kg weight at arm-length) but it masses 540Kg and has a rated power consumption of 5KW. (Three-phase power, of course) And that's just the arm! The NX100-FM controller it's specced with masses another 120Kg.

Mobile robots hate weight. Cutting weight forces a lot of other compromises, in speed, power, and cost.

What about Atlas ?

http://www.bostondynamics.com/robot_Atlas.html

EDIT: Also for the Motoman, is it possible it needs so much power because of how fast it can move that 15 kg mass around ? There's no beating conservation of energy.

> Atlas is powered from an off-board, electric power supply via a flexible tether.

There's your problem right here. The video demo shows it walking freely, but I bet it can only do that for short stints.

If you studied robotics, you'd have learned that some walking/running problems are easier to solve when you have an engine with "unlimited" short-term torque. Low-powered electric actuators are very bad, hence BD uses gasoline engines to provide that short spike necessary for some differential equations to have a nice solution (e.g. one without power-pumping or without needing a few cycles around to reach your desired state).

I'm okay if my household robot is stuck in my household, and basically consists of very long arms.

Jonathan Hurst, the inventor of the ATRIAS robot, has argued that we have the actuators necessary and that efficiency solves some of the problems with batteries[0]. A brushless motor hooked up to a big reduction gearbox can have a pretty high torque density. If we design our robots to be light and design them such that they don't throw away energy with every step, then they can go further with current batteries.

[0]http://tedxtalks.ted.com/video/Walking-and-Running-Bio-Inspi...

Reply on news.ycombinator.com