Will robot hands one day replace vacuum grippers in the warehouse? No chance.
First, dispelling a myth
There is an unfortunate myth, that vacuum gripping requires a seal. Some vacuum systems do require a seal, but not all. In fact, you probably have some in your own home — your vacuum cleaner, DustBuster, Roomba, and Shop-Vac all pick things up without a seal. True, the usual vacuum cleaner is tuned for picking up tiny things, but if you have played with a Shop-Vac you know it can do more. In fact, the winners of the first Amazon Picking Challenge used a Shop-Vac-style vacuum source.
Looking at robot hands
Investors are spending billions of dollars to develop humanoid robots, including anthropomorphic robot hands. Believers say the robot hands will do everything human hands can do. But this is an ambitious goal. For example, the number one wear point will be the fingertip skin, with thousands of embedded touch sensors. In human hands this problem is managed by constant self-repair, renewing the skin as it wears, from the inside. Will the robots be repairing themselves from the inside? Doubtful.
Another thing. Scientists and roboticists are still trying to catalog the vast number of human hand behaviors. As an example, in a recent blog post (The Inner Robot) I discussed the ulnar grasp, which we humans use occasionally, but which has gone almost unnoticed in the scientific literature. Scientists have been producing grasp taxonomies for well over 50 years, but they missed the ulnar grasp and many other behaviors. How much longer will it take to achieve a thorough understanding of human manipulative behaviors? The limitations of our knowledge are well-known in the robotics research community. There are regular discussions of the need for metrics and benchmarks. Some have been offered, but nobody would claim to have captured or delineated the limits of human manipulation skills. Keep that in mind when you watch the dazzling online videos. As impressive as they are, we still have a long way to go — longer than anyone can say.
In short, reproducing the capabilities of the human hand is a grand research project, to be supported and celebrated. But it is still a research project, and nobody knows when or if it might lead to practical robotic hands.
Comparing
Nonetheless, let’s just concede the point and move on to the next question. Let’s suppose that robot hands were available and as capable as human hands. Would they replace vacuum grippers in the warehouse? No.
The first thing to look at is performance. Vacuum grippers have a dramatic performance advantage in logistics: they can grip an object using only one contact patch. Imagine a card lying flat on the bottom of a bin, in the corner. A hand, whether human or robotic, could fiddle around for quite a while, trying to get two fingers on opposite sides of the card. The vacuum gripper only needs to touch it, and it has it. Or suppose you are taking the first box from a densely packed case. The vacuum gripper can touch any box in the top layer, and it has it. The hand will need to force fingers into the gaps, without damaging the product. That takes time, in the best case.
Now you might be thinking of some items where the hand wins. That’s fine. A warehouse can have several different picking systems. If there were items a vacuum gripper could not handle, you could steer them to the other system, which could be a human being, or, conceivably, an anthropomorphic robot. So yes, hands might find a place in the warehouse, but no, they will not displace vacuum grippers.
Beyond performance, the comparison is quite easy. If the robot hand designs copy the human hand, they will have:
- Five fingers,
- Three joints per finger for a total of 15,
- Four expensive servomotors per finger for a total of 20,
- An additional joint and motor in the palm,
- Tens of thousands of tactile sensors.
A typical vacuum gripper has
- One “finger,”
- No joints,
- One or two motors, typically off board,
- One pressure sensor.
As a result, the vacuum gripper will be simpler, lighter, more rugged, less bulky, faster, and less expensive. These differences are not minor. In some respects, the vacuum gripper wins by an order of magnitude. Even if some researchers dialed back their ambitions and simplified their designs, they could not match the elegant simplicity of a vacuum gripper.
The bellows
Vacuum grippers often come with a bellows — a springy accordion-like cylinder just above the cup. The bellows flexes to make good contact, even if the object isn’t where expected. When vacuum is achieved, the bellows contracts, lifting the grasped item.
The bellows is so simple that it is under-appreciated. Obviously, it helps to deal with placement error, and its lifting behavior can be useful. But there is more. First, when fully contracted, it stiffens. That’s important. With the object stiffly connected to the robot, the robot can move quickly without the object flopping around.
And another thing. Let’s say the goal is to pick a flat horizontal thing, but the thing is resting on another object with one end tilted up. The bellows bends to get a good grasp, and then it retracts and stiffens, but it also returns to horizontal. The bellows does not just accommodate the error, it corrects it.
Vacuum grippers are the wheels of manipulation
There is an interesting parallel between vacuum grippers and wheels. For centuries, humanity has been using wheels to do the job of legs: supporting and transporting stuff. But now, there is an intense effort to develop robot legs. They are making excellent progress.
Will legs replace wheels? In some specific areas, they might. Legged humanoids have already appeared in warehouses. But will legs replace wheels broadly? No chance. Wheels are one of the most celebrated inventions of humanity, and for good reason. They are simple, light, rugged, and inexpensive, like vacuum grippers. They require a minimum of motors and sensors, like vacuum grippers. The case for wheels is remarkably similar to the case for vacuum grippers.
Wheels will be with us forever. So will vacuum grippers.
Notes and details
Note that I am employed by Berkshire Grey. This post was adapted from a Berkshire Grey blog post.
See a summary of the first Amazon picking challenge here: Correll, N. et al., “Analysis and observations from the first Amazon picking challenge.” IEEE Transactions on Automation Science and Engineering, vol. 15, no. 1, pp. 172-188, Jan 2018. The winning team, RBO from the Technische Universität Berlin, relied exclusively on vacuum gripping.
Am I obsessed by vacuum gripping? My interest was strong enough that Chris Atkeson produced this funny video: Paradigm Shift: Vacuum Cups.
I tracked down the original bellows vacuum gripper patent here: Jackson, John. Article lifting device US 2,798,757, awarded 1957. Jackson’s description includes the stiffening and error-correcting features of the bellows. Smart guy!
“Vacuum grippers are the wheels of manipulation.” I love this. Of course, two-jaw grippers might complain. Perhaps they are the tank treads of manipulation?
But perhaps we should compare vacuum grippers to tools (e.g., screwdriver, hammer, scissors, knives, etc. etc.; those will all be with us forever, too) instead of to dexterous hands. Glass handlers, for example, use suction gripper tools all the time. In that view, dexterous hands are versatile tool changing systems; so versatile, in fact, that they sometimes don’t need tools, and in fact certain tasks that can be performed by hands may not be possible with any known tool.
I agree with the main point of this blog post: just as legged mobile robot folks need to justify their mobility platforms relative to wheeled or treaded platforms, based on the variety, relevance, and value of tasks they can perform, dexterous hand folks need to justify complex designs relative to simple grippers/tools based on the variety, relevance, and value of tasks they can perform. And, as you say, even human grasp types are not fully characterized, much less the much larger variety of in-hand manipulations (where object motions are controlled relative to the palm and fingers), and much much less the still larger class of bimanual, trimanual, and quadrimanual in-hand manipulations. Lots of work for us to do. Exciting times!
In line with the topic of the post, my reply was very manipulation-centric, but we could also consider exploration and active sensing taks that hands/grippers/tools could be assigned; for example, sensing texture, hardness, temperature, etc.
Thanks Kevin!
Tools! So much to say, but I cannot seem to develop a coherent line of thought.
Today I used AI with over a trillion parameters, trained on a petabyte of data, to help me find a synonym. Not a very efficient way of doing things, but it sure was easy!
Vacuum grippers suck! 😉
Enjoyed reading this article. I am glad that you call out the hype and the force-fitting these days of legged robots for literally all use cases.
Are you of the opinion that two jaw grippers and vacuum grippers suffice for the general purpose robots in the household?
Thanks! Re the sucking pun, I tried and tried to squeeze it into the blog post, and couldn’t. So thanks.
Re general purpose household robots. I am not sure I believe in such robots at all. As an engineer, or a wannabe-engineer, I am disturbed by the idea of working on such a poorly stated problem. We cannot figure out how to describe the problem, so instead we try to copy an existing solution.
But still, what do I expect to see in the kitchen, someday? Perhaps we will have some special-purpose machines that emulate some of the humans’ most impressive performances on common tasks? And then some other machine that can sometimes take care of less common tasks? Like when you accidentally drop a fork behind the stove? But, actually, nothing I dream up seems plausible.
Thank you Matt for this insightful and eloquent article. It truly sucks. When I was doing my recycling work, I was amazed by the number of patents dedicating to suction cups. I found this article, and the following comments, to be educational for us roboticists and those who care to invest in robotics. I do have one small thing to add. One of the justifications for developing and research robotics hands is that hands are biomimetic. Of course, pursue of robotic hands as a research endeavor should be celebrated and supported, as Matt put it. However, suction is also biomimetic too. Lizards have tongues that fire out and then suck in a variety of yummy treats. Octopuses have suction cups on their arms. I need to learn more about the bio-mechanisms of sucking as well.
I would have said that rather than being error-corrective, using a suction cup is uncertainty-tolerant, i.e. using one could be considered an example of a manipulation funnel, making it wholly consistent with your research program 🙂