Isn't making a robot that looks like a human to replace a human's job a bit like making a mechanical horse to fulfill our transportation needs?
Having worked with robotics for years I can say the amount of setup work that goes into installing fully-functional hardware and software and getting a robotic process running smoothly is enormous. The basic idea is that everything, the robot, all the hardware, all the firmware/software, the end-effector, the workpieces, the sensors, etc. is rigidly-defined and over-spec'd so that with all the tolerance stackup and after all the integration and process debugging work you set it and DON'T CHANGE IT for as long as possible. The chaos that ensues from one little component changing it's behavior can be enormous.
The notion of "smart" robots that you can just slap down or that can just handle all sorts of unknowns and adjust themselves to changes to me always seemed like a really, really big challenge, maybe not as challenging as a driverless car but definitely more of a "general AI" problem.
I'm sure someone has coined the term but there must be some kind of "uncanny valley" of intelligence: a little intelligence (e.g. the PID controllers that actually run robots) is great, a lot of intelligence (fully-blown general AI) is great (if you can get it), but what's in the middle may not be worth the while. Getting the answer correctly 99% of the time doesn't work if you need 99.9% success rate.
From an investment standpoint I would be looking for companies with a REALLY specific well-defined problem that "medium AI" could solve rather than someone who's claiming to take medium AI and apply it vaguely/generally.
I guess that's my take away from this: work on specifying the problem before you work on the solution.
I see current AI to be like donkeys, and the middling AI you speak of as chimps. There's a reason we domesticated donkeys and not chimps.
The autopilot feature of Teslas is a lot like a donkey. It mostly handles itself but is stupid and needs a lot of monitoring. Using autopilot feels a lot like sitting on a cart and pulling the ropes on the donkey every once in a while.
"Medium AI" is definitely useful, as long as you have the correct interfacing and apply it to the correct problems.
Obviously you can't just slap it onto an AI-complete problem and have medium AI perform well enough to ship.
We are probably roughly in agreement here, but I disagree that in reality there's a significant uncanny valley effect. It's just hard for those not in the field to intuit about the capabilities of medium AI.
It's not that the concept is a failure. It's that Rethink, the company, is a failure. Universal Robotics in Denmark is doing something very similar, but they have over 25,000 installed robots and distributors in 50 countries. They also have a robot arm that is a nice piece of mechanical engineering, and retains its precision over a very large number of cycles.
The coming thing is robots as a service, paid by the hour. Hirebotics offers that. They own the robots. They set them up. They fix them. You pay them for each hour of use, like an employee.
Vision for robots today works well for semi-structured situations. If you know what the parts look like, and can see the parts and target location reasonably well, you can probably get a vision system to guide a robot arm to put things where they're supposed to go. The fixturing no longer has to be so rigid that a robot can do the job blind. The systems that do this are often just convolving a stored image of the target against a camera image of the area that contains the target. This is 1970s technology, but we have enough cheap compute power now to make it work fast.
I have been working with the Sawyer arm for the past few months and it's a really nice piece of hardware - far better than the Universal Robots arms that it competes with, while being a similar price. The software is also far ahead of the crappy 2000's era interface of the UR arms, and they have a nice ROS-integrated SDK for people who want to go farther with it. I was really disappointed to see them shut down as we were hoping to buy several more in the next few months.
I was an intern at Rethink (back then, Heartland Robotics) 2010~2011, it was an exciting place to work at. There were ~10 employees when I joined, and ~40 when I left.
One of the early product mistaked they made was focusing way too much on low-cost. They compromised many necessary performance specs (repeatability, speed etc.) just to be able to make the robot out of plastic. They eventually pivoted to their second generation robot, Sawyer, which was made out of casted metal components. This gave the robot much better performance but I guess that it was too late.
I would really love to hear why performance specs you mention are still necessary for robotics applications. Repeatability and speed certainly seem useful industrial robotics applications, but low cost also seems critical to enabling broader use of robots. I am particularly interested in the possibility that modern algorithms (e.g. deep reinforcement learning) can compensate for not having completely repeatable behaviors from a robot by using sensory feedback to create behaviors that reliable despite their variability. Additionally, materials with higher elasticity (e.g. plastics instead of metal) might allow some previous impossible efficiencies, e.g. by storing potential energy in the links and then releasing it at the right time. Is it just too soon for these sorts of approaches?
Compromise on stiffness and rigidity in a servo system means the thing jitters about like a squirrel on speed. Robots need to be smooth, and there's precious little you can do in your controls to compensate for system instability. Precision comes along almost for free once you have that rigidity.
They didn't get the basics right. Special sauce on top needs to have a solid foundation.
And that's why I have an Aubo robot in my conference room today, and why I was able to tune in a 250kg payload 1995-vintage ABB on Tuesday, while these guys are shutting down.
A robot is basically a programmable jig. If a robot is floppy like a human, and needs a traditional jig for precision work, then the jig is inflexible and you might as well just build a single-purpose machine to do that thing.
traditional robots need to be smooth. I doubt commoditized robots will be based on traditional servos with traditional pid control, there is plenty of room for algorithm and control improvement, not too mention the huge strides being made in soft actuators. Like a post below you said, humans are able to do it, so its a question of how to cheaply artificially do that. Just because something doesn't exist yet, doesn't mean its impossible, that's the whole point of research. As an example, the group I was a part of in school made specialized control algorithms for 3D printers that can compensate for the flexibility and dynamics of cheap/thin actuators, which nearly doubled printing speed while retaining accuracy
Or tune in a 250kg payload on such a robot - I can imagine the motion programming might have to be tweaked to swing right. I don't know what such a heave payload might be - maybe some heavy part to be welded.
By 'tune in' he means calibrate an end effector to be mounted on the robot. This involves TCP calibration (tool center point) and specifying the centre of mass and moments of inertia. KUKA robots, for example, allow you to determine the TCP by manually posing the robot so that the tip of the end effector (tool) is at a fixed point. This is then repeated three more times from different poses (but the same tool point) so that the robot can solve for the 3 values that determine the tool's XYZ offset from the flange (mounting point at the end of the robot). The robot controller then uses this data when calculating inverse kinematics and supplying power to the axes.
>> I am particularly interested in the possibility that modern algorithms (e.g. deep reinforcement learning) can compensate for not having completely repeatable behaviors from a robot by using sensory feedback to create behaviors that reliable despite their variability.
Deep Reinforcement Learning will definitely make a huge difference in robotics. Already, ~50 years after the first implementations of RL algorithms [1] robotic systems trained with (deep) RL have now become so proficient that a robotic hand has learned to manipulate a cube [2]. Always the same cube (in terms of size, weight, etc) but a cube, nonetheless! Full-blown general object manipulation cannot be more than a hundred years away, or two!
__________________
[1] I'm counting from Donald Michie's MENACE, a tick-tack-toe playing machine originally implemented in matchboxes, because of a dearth of computing hardware of sufficient power- see Rodney Brooks' excellent write-up:
Note that it's always the same cube. If they could train it to manipulate anything else, you can bet they'd be showing it off (and probably publishing in Nature or something).
Plastic vs metal has nothing to do with it. A series elastic actuator is going to introduce error far larger than a standard harmonic gear drive motor whatever you make it out of.
Two bridges into a down round on their series E, these guys have been bleeding money for years. The VCs gave them ten years to find a sustainable path and they just never took off. This is just how it goes. It’s probably for the best, and 10 years in any venture is a fair shot.
I work in robotics field and I always thought that collaborative robots were cool in trade show booths but beyond that they had limited use. I've not seen many of these things used in our industry (semiconductor) or any other friends who work in various industries.
ABB lead the campaign on these as if they are going to revolutionize manufacturing lines where humans & machines will work together. That never happened just like 3D printing never shook the world as it promised it would.
I used to think so but it's much closer than you think.
My side project we're using a Prusa Mk3 for an end use part and the quality and cost is amazing. A lot of other people are doing the same. There are so many types of material too now.
I work on 3D printing and I can tell you - 3D printing is profoundly impacting a variety of industries. It’s been amazing to see the growth and development in the past 3 years.
Yeah, I think the thing with 3D printing is it's a powerful industrial technology, but the idea that every home would want a low cost 3D printer was badly wrong.
It's just not good enough yet. It just can't make enough products using a variety of materials, fast enough and with little user involvement. When a turn key solution is available where the average person just clicks to "print" a product, that equation changes.
I feel like this is a technology that isn't a consumer product until suddenly it is.
The problem here does not seem to be with collaborative robots, as much as execution, with Universal Robots outcompeting them. Honestly the friendly face UI seems to be gimmicky. I suspect more practical features were under prioritized.
We have URs in a warehouses doing piece picking. They're not the right solution for very structured environments like a semiconductor plant but for medium structured environments like a warehouse they can make a lot of sense.
And frankly collaborative arms are just so much easier to develop on that it might be worth it to use them for that even if your final product is going to use a full industrial arm. You're going to make a mistake in your collision checking routine at some point.
Kind of sad to see them shut down. Been following them since the Baxter robot came out. Unfortunately, the current market just does not have a high demand for highly flexible robots. Most of the fully automated lines are highly specialized lines. Company invest hundreds of millions, with the expectation of maintaining the line for years. It's usually simpler and cheaper to use standard actuators with multiple stations rather than a complex robotic arm. However, the couple of applications from Universal Robotics and Kuka is pretty cool to see.
Sad to see them go. I spent just under a year there ~2 years ago doing embedded systems, right before some big layoffs, and I really had a great time there.
It's a great group of people, and I wish them all the best. Judging by the amount of recruiter calls and emails I got today after the news broke (apparently working off my old resume!), I think these folks will land on their feet!
The trouble with the thing was that about all it could learn was how to pick up something from one place and put it somewhere else. That's useful, but there are already lots of other robots which do that, including ones with fast vision systems.
Baxter robot, finding and picking up simple un-oriented objects and moving them.[1] Slowly.
Festo robot, finding picking up simple un-oriented objects and moving them.[2] Fast.
I sold Baxter and Sawyer robots. They were shit. The creators had no idea about market fit. You’re selling this thing to controls and automation engineers who are familiar with programming on PCs and pendants. ReThink had this ridiculous face for “emotional feedback” that forced you to also do all your programming there. It was clunky and kinda cheap (Sawyer was upgraded, and more solid). Baxter was slow with terrible repeatability. And it did not come with native ports for external controls or communication. There was no way to integrate with an existing automated system, no digital IOs or Ethernet. You had to get a third party distributor to manufacture at their panel shop and jury rig something together for extra cost.
Not to mention that UX was terrible and clunky and not intuitive. It’s suppose to be easy to use, easy to program, that anyone can just role it up to an existing operators position and teach it in five minutes and you’re up and running. Wrong.
As mentioned, the only interface was the obnoxious swivel screen, and the only programming controls were... dials with a push button, one located on each arm and one on the back. Yes, you had to scroll through menus and across pages to select and modify programs in endlessly confusing menus that literally made you regret ever getting involved with the company.
Yes, it had some nice collaborative safety features built in, some patented titanium s joints that flexed for safety like tendons yet maintained nice accuracy.
But no one cares about your fancey safety features if it’s nearly impossible to use, and fails to seamlessly integrate into existing systems.
This is a classic mistake. Build something epically fancy, but totally worthless to 90% of the people that would use it, because it just doesn’t play well. Learning curve is too high. It’s not plug and play/ fails to understand industry compatibility requirements, or it’s too niche.
Because I sold these things, I quickly learned they were terrible.
However, Universal Robots were on the right track from the get go. They were coming from the industrial automation world, they understood market fit. They made a device that was easy to use, intuitive, that leveraged existing programming methods and incorporated some collaborative programming technology in tandem. Maybe their gear mechanisms weren’t as safe, but they passed all the requirements, so what did it matter in the end? Fancy safety joints were NOT the most important selling point for collaborative robots. Reducing downtime to increase output and profits is, and that means easy integration and quick programming. Not to mention UR was just a robot arm, and not this wacky monstrosity of a machine and wheeled based (which was optional for ReThink but not really). And UR was super compatible. It was basically like conventional single arm robot with servos slowed down and given fancy torque feedback algorithms and some sensors for safety and tracking.
The guys who started ReThink were out of their depth. Just some academics who had some success making commercial robots who expected an entire industry to conform to their fanciful dreams of robots with dopey emoting faces working along side their human comrads.
Just pissed money away.
And the thing is, they were one of the first, and positioned to be one of the best. Collaborative robotics is a massively growing field as humans and machines work in ever closer proximity.
But ReThink totally choked. Had the wrong design engineers, with no industry experience. They failed to understand industry needs and existing automation culture and prevailing systems infrastructure.
This company is a classic case study on how to completely botch the opportunity of a lifetime because of failing to understand existing market fit.
Interesting that one of the top articles today is about finding the Apple II LOGO source code. One of the authors of the code is Patrick G. Sobalvarro, who also happened to be president of Rethink.
>Rethink will now be selling off its intellectual property and patent portfolio...
Hopefully they sell to someone who intends to put them into practice and not some troll so these patents don't become another toxic ip spill the industry will spend the next 20 years navigating around.
Aaaw. That's Rodney Brooks' company! I warmed up to the crazy old coot, thanks to his recent series of articles on AI and machine learning [1]. It's very unfortunate that his company tanked.
That's a treasure trove of knowledge on AI with a generous dose of personal, um, perspective, but from a gentleman who has a very, very long career in the field. Do read his stuff if you fancy yourself "knowledgeable about AI".
Wow they had a lot of positions open before they closed, was always tempted to apply. Sad for the Boston robotics scene and the larger commercial robotics industry.
How do they compare to Boston Robotics. My understanding they were pursuing different use cases. Former is more about doing repetitive work and latter is more to do movement and carrying stuff. Although Rethink's robot look much primitive to Boston's robots, however may be what they're trying to achieve is much more complicated?
Boston Dynamics is focused on building free-moving legged robots.
ReThink was focused on building industrial robots (think robot arms) that can easily be trained/retrained to do a wide variety of tasks and that can co-exist with humans (torque sensors tell them when they’ve hit something unexpected and tell the arm to stop before it crushes you). The idea was that smaller manufacturing settings could use them side by side with human workers, and that a single arm could be retrained frequently as tasks change over time. This is compared to traditional industrial robots which have to be completely cordoned off from humans lest they run into them, and which require extensive programming to execute a new task.
Sad news. It's one of those things that will eventually take off -- and not too far in the future. So to me it just goes to show that today's markets are very sub-optimal, overly favoring short-term returns.
Having worked with robotics for years I can say the amount of setup work that goes into installing fully-functional hardware and software and getting a robotic process running smoothly is enormous. The basic idea is that everything, the robot, all the hardware, all the firmware/software, the end-effector, the workpieces, the sensors, etc. is rigidly-defined and over-spec'd so that with all the tolerance stackup and after all the integration and process debugging work you set it and DON'T CHANGE IT for as long as possible. The chaos that ensues from one little component changing it's behavior can be enormous.
The notion of "smart" robots that you can just slap down or that can just handle all sorts of unknowns and adjust themselves to changes to me always seemed like a really, really big challenge, maybe not as challenging as a driverless car but definitely more of a "general AI" problem.
I'm sure someone has coined the term but there must be some kind of "uncanny valley" of intelligence: a little intelligence (e.g. the PID controllers that actually run robots) is great, a lot of intelligence (fully-blown general AI) is great (if you can get it), but what's in the middle may not be worth the while. Getting the answer correctly 99% of the time doesn't work if you need 99.9% success rate.
From an investment standpoint I would be looking for companies with a REALLY specific well-defined problem that "medium AI" could solve rather than someone who's claiming to take medium AI and apply it vaguely/generally.
I guess that's my take away from this: work on specifying the problem before you work on the solution.