E76: Mitchell Weiss, CTO at Seegrid – Interview

November 22, 2016


This interview is with Mitchell Weiss, who is the CTO of Seegrid. Seegrid develops and sells autonomous robots that navigate manufacturing and distribution centers. They help companies move products around without a human operator. Here are videos showing their robots in action:


Seegrid has been working on robots for a long time – since 2003. Mitchell joined one year later in 2004.

Mitchell has a wonderful background in robotics and automation including a degree from MIT. He has 23 patents, and he serves as an expert witness in IP litigation. He also has experience with product design, production and sales.

I invited Mitchell on the show to learn more about what he’s doing now at Seegrid and what he thinks about the future of robotics and automation.

Some other things we talk about:

-Seegrid thinks about autonomous driving and safety differently than most car companies. What advice would you have for self-driving companies?
-Where do you want to take Seegrid’s tech in the future? It’s not around self-driving as much.
-What’s the process to map out a distribution center for a new client?
-How does your tech work?
-How’s your safety record? Pretty shockingly good.


David Kruse: Today we get to have Mitchell Weiss with us. And Mitchell is the CTO of Seegrid. Seegrid develops and sells robots that navigate manufacturing and distribution centers they help stuff move around essentially without a human operator. You can check out their videos – they’re pretty fun to watch. They’ve been doing this for a long time. So, Mitchell has a wonderful background around robotics and automation which he can tell us about. He has 23 patents and is an expert witness in IP litigation. And he has experience with product design, production and selling. So, I invited Mitchell on the show to learn more about what he’s doing at Seegrid and what he thinks about the future of robotics and automation. So, Mitchell, thanks for coming on the show today.

Mitchell Weiss: Thanks Dave and its good to be here.

David Kruse: So before we get into the, kind of what you are up to now, can you give us a little bit of an overview on your background?

Mitchell Weiss: Yeah, it’s a fun story I could tell. So when I was 16 years old, which is a long, long time ago in high school, I was reading Isaac Asimov’s I Robot and one of my classmates dropped on my desk a flyer for the local science fair and said, ‘you should enter something in the science fair.’ I said, ‘yeah, I should’ and I built a robot. So at sixteen I started building a robot. And then when I went off to college I decided I wanted to stay in the robot biz. It was suggested to me I try this funny American school called MIT and I grew up in Canada, so I went to MIT and I wanted to study robots and they thought I was out of my mind because I wound never find a job and I studied what I needed to learn and I’m now in the robot biz.

David Kruse: Interesting. So back then MIT was not into robotics?

Mitchell Weiss: There were two classes at MIT that covered robotics back in 1977. There was one called making machines see and feel, which later became called robot manipulation and vision. And there was another one called, I think it was robotics or something to that effect, and that course wasn’t even offered by the time I was able to take it.

David Kruse: That’s crazy, but they have a few more now, probably have a…

Mitchell Weiss: Yeah. I think so.

David Kruse: 50. And weren’t you a lecturer at one point at MIT before Seegrid.

Mitchell Weiss: Yeah, yeah. Well I worked as a part time lecturer at PIT and MIT, depending on where I was living at the time, and I have lectured at MIT in the machine design course, the class where the kids make little machines that compete against each other.

David Kruse: Cool! And what did you do after graduating from MIT. What were some of the companies you worked for or projects you worked on?

Mitchell Weiss: The first company I went to work for Unimation, which was – and so your listeners should know the first robot company on the planet. It was founded in 1959 by Jo Engelberger and oh! Golly, I forget his name – sorry, and George Devol. George Devol invented the Unimate and Jo Engelberger was the entrepreneur. In ’61 they installed the first robot and I went to work there in ‘79 when I got out of school and I was the first Application of Engineer on the PUMA robot, which was their first electric computer controlled robot. Then I left Unimation and started a company called the United States Robots in 1980 with another Ex-Unimation employee. And at the time we were the fifth robot company in the U.S.. In 1981 there were 20 of them, it was crazy. So around 1980 that the industry really took off and rode that way, sold that company in ’82, started a material handing company in ’85 that made material handling equipments for the semiconductor industry and we sold that to PRI in Boston in ’93 and we became the number one supplier of automation equipment to the semiconductor industry worldwide.

David Kruse: Wow!

Mitchell Weiss: Then after that I joint Seegrid in 2004.

David Kruse: Got you, okay and out of all those experiences, what was a memorable project, whether good or bad that you – what you learnt from it?

Mitchell Weiss: Well, everyone is a learning opportunity and there were really two big things that got learnt. One was the robot business was really hard, at least it’s really hard to win big at. It’s not a unicorn kind of business like coming up with a web app where you can do it in your garage by yourself, because you got to bring all this mechanical, electrical and probably most importantly domain knowledge about the customers problem to the business and so it’s hard to do. And it’s hard thing to get abduction on, because it’s not a consumer product, it’s a business product and so the users are industrial users who have to change their process and flow to use this stuff; so that’s hard to do. It’s a little hard, but that’s the biggest learning. Probably the – I mean every time I come up with something new or invent something new it’s kind of fun and exciting, but I guess the first robot we designed and built at U.S. robots in 1980, that’s sort of a benchmark.

David Kruse: What do those do?

Mitchell Weiss: Especially the first time you see it work.

David Kruse: What do those robots do?

Mitchell Weiss: It was industrial manipulator. It was a human scale robot for loading and unloading machines and doing precession assembly work and stuff. So we had done some of the applications we had. We were moving silicon wafers around, we were stuffing circuit boards, we were doing mechanical assemblies, building smoke detectors, things like that.

David Kruse: And how is robotic design and development now compared to back in 1980 as far as ease of development?

Mitchell Weiss: Yeah, well it’s obviously a much easier environment to develop in now. We at U.S. Robots we had to design our own motion controllers, our own power amplifiers for the motors. We had to work with the manufactures and the motors and the gear trains and things to debug their manufacturing processes. It just wasn’t a mature market yet. One of the products I worked on before that, some of our job was on the space shuttle and working on the arm on the space shuttle. The space shuttle flew with core memory in the computer. For your readers who don’t remember core memory, they are little, they are like donuts with wires going though them and each one represented a bit. So the memory that threw them on the space shuttle in the 70s was essentially mechanical.

David Kruse: Interesting.

Mitchell Weiss: We couldn’t fly semiconductors into space, because they couldn’t handle the radiation.

David Kruse: Well…

Mitchell Weiss: So it’s a lot easier now.

David Kruse: Yes, and we’ve had some people on who are trying to help make it easier. But yes, so we – the whole podcast could be on your background; you’ve done a lot of interesting things, but let’s talk about Seegrid and what prompted you to join Seegrid and when was Seegrid started and when did you join?

Mitchell Weiss: Seegrid started in 2003, and it’s based in Pittsburgh, Pennsylvania. It’s a spinoff from Carnegie Mellon. I joined in early 2004 and what prompted me to join was a thing that Seegrid does. So while what we deliver to our customers are self driving vehicles for use in factories and distribution centers and warehouses. What we are based on is the software, and the software approach that Hans Moravec, who is one of our Founders and was formally the Director of the Mobile Robot Lab at CMU, who built his first robot before I built my first robot, so he has been in the business 10 years longer than I have. Our navigation technology and our vision technology is all based on statistics. So it has a mathematical basis as opposed to juristic, which is just algorithmic and when I came to visit in Pittsburgh with Hans and he explained to me how his system works based on statistics, I got excited. You know that’s kind of super nerdy geeky detail, but artificial intelligence work for years was wrapped around clever tricks, meeting juristics and was not succeeding and here Hans was able to explain to me that he had this process for doing navigation and vision that was based on statistical map and I said, ‘well, that’s pretty convincing. You should make that work.’

David Kruse: And what’s the – can you give an example of what’s the difference between using the algorithm approach versus statistics?

Mitchell Weiss: Sure, the way our system works is it captures a lot of images in the environment. It processes those images and it finds a lot of features in the environment. So if you look around the room you are in, the edge of the light switch, the edge of the door frame, those are a bunch of specific features. They don’t have a lot of meaning like a door or a sealing pile, but you might have the edge of the tile, the edge of the ceiling. So if you are using your juristic and you are trying to navigate, you might tell the system, ‘okay, I have computed that this thing is a door and you are three feet from the door.’ If you are using statistics, we don’t care that it’s a door, we just care that there is a bunch of obvious features around it and I’m not going to try and interpret what they are. But if I grab enough of those features, so the door frame is the corners of the room, stains on the carpet, ceiling tiles, light fixtures, you know just a lot of random edging in the room and I don’t figure out what they are. I just have to measure my distances and identify where I am with respect to that pattern of points which is – I presume it’s a bunch of math. So I now have a system that’s very robust because it’s based on large amounts of data, and I haven’t parched it into things that are important to humans, because human don’t work that way. I mean, we don’t process lots of statically data to figure out where we are. We say, ‘oh! We’re near the door.’ So that’s the big difference and that is – it’s very computationally intensive, but it’s very good for what computers are good at.

David Kruse: Interesting and let’s see, I have a lot more questions about that, but first can you just give a kind of a brief overview on Seegrid, if you know, you know the number of employees and the products that are you offer?

Mitchell Weiss: Yeah, I got it.

David Kruse: Yeah.

Mitchell Weiss: All right, so Seegrid is in Pittsburg. We have about 100 people working here. We’ve got people in Europe and scattered around the U.S. in the sales and support sides and stuff. It’s about 100 people in the company. Starting in 2003 our core business is what we call 3D Perception, which is the software side of our business and what we do with that is we currently take things like industrial toe tractors or industrial pallet trucks or forklift kind of trucks and we add components to those, which include our home grid stereo cameras, we make our own stereo ranging cameras, computing platforms, safety systems and we turn those industrial trucks into driverless trucks. So it’s actually a truck that you can get on and drive and move around just like a manual version, but you can also step off it and send it somewhere on its own. The reason we did it that way was we didn’t want it to be in the truck and vehicle making business. That’s a commodity business where other people would do that really well. What we wanted to do is be in the – give the truck brains business. So we have well, our stereo ranging cameras and the ability to build maps that the vehicles navigate around, and then the tools we need on the trucks, so that people can easily use it. So our system is a teach repeat system. We call it ‘Walk Through Then Work.’ You move the truck along the path through the facility and we can store up to 15, 20 miles worth of paths on a truck and once you have driven around once, you can repeat that. So it’s a very simple show it the job once and then it does the job again. And the vehicles navigate within a couple of inches of the desired path. When you have a lot of our vehicles in the building and you need to coordinate their activity. So for that we have another product that’s called supervisor. So all the vehicles navigate on their own, but they coordinate their activity by talking over the facility WiFi back to supervisor and supervisor coordinates where they are going, when one truck can go through an intersection and another truck has to stop and we do all of that without adding any infrastructure to the customer. So one of the big advantages of using our vision systems and the way we train the trucks, when we come to an intersection, we just tell the truck this is an intersection. I don’t have to stick a traffic light in or a gate or a switch or a doorway. The truck knows this location in space in an intersection. I will ask for permission and go through it, before I go though it the next time. So we literally can take one of our vehicles to a customer site, turn it on, run a route, make be the route is a 100 meters, a 1,000 meters long and repeat that route within minutes of showing that they are out and in fact that’s how we demo to our customers.

David Kruse: So you don’t have to map anything, I mean the mapping is just drive around the truck and they will memories such as positions.

Mitchell Weiss: Right, that’s correct.

David Kruse: So how does it memories its position? What is it is using the sensors on board, the WiFi signals, or how does it figure out the whole thing?

Mitchell Weiss: Well, it’s all based on the vision system. So on each truck, on each truck there is five stereo cameras, so there is 10 imagers all together arranged with five stereo cameras and with the stereo camera, when it looks at your left eye and your right eye blinking one eye and you will see things shifting right to left because of the parallax. So the difference in the two images is called the disparity. So based on the disparity between points and the image, you know how far away those points are. So imagine you don’t have two eyes on the front of your head, which have two on the front, two on the back, two on each side and two on the top. That’s what we have on our vehicles and there is a thing called the visual guidance unit and has these five stereo cameras on it and instead of looking all around the truck at 360 degrees and above and below the truck, almost at 360, can’t see through its own feet. As the truck is driving, as we’re training it, it’s taking pictures, it’s talking stereo pictures and then it processes the data in those stereo pictures and for every snapshot that it takes, its finds a 1,000 to 2,000 features in each snapshot the five camera presents. It then builds those features into the 3D map. That’s where all the statistics come in, because we built them into a map and I think it’s called an evidence grid which is actually a statistical probabilistic work presentation of the space. So again, we are not trying to show a picture of every column and being in the building. We are just showing a random sampling of a lot of important points in the building. So once you’ve built that whole 3D map, you take the vehicle to a start position. You send it to a destination by pushing a button on the user interface and then as its driving to that destination it’s continuing to take these snapshots in all these five stereo cameras and its comparing its position to where it is in the map. So it has an idea roughly where it is. It looks at those thousands of points in the images. It compares the fits of those points to the map and it says, ‘oh! yeah I’m within 20 millimeter of where I should be,’ and that’s how it works.

David Kruse: What if those points – I’m sorry.

Mitchell Weiss: And that the – go ahead.

David Kruse: I was going to say, what if those points on the map change. You know like they go through and a barrel was there before and now it’s not, how does it not get confused?

Mitchell Weiss: There is the big one of statistics. So because we are capturing points in 360% and we are capturing thousands of them on each glimpse, if you move a bunch of stuff around them in the building, there is still enough stuff to tell you where your best fit is and your location you. If you are using a LIDAR based approach, where someone is using a 2D LIDAR scanner or even a three or four, five beams scanner, they are only able to collect dozens of points. And so if you move a pallet of goods, they lose a lot of their reference position and now they can’t do a good analysis of where they are. So that’s the advantage of vision. We capture so much data in a single electronic event, right, that if we apply a lot of statistical processing to that data, we have a very robust understanding of where we are, even if the environment is changing and in all the facilities where the environments changing. But it’s not chaining enough that it looks like a different planet.

David Kruse: So, that’s pretty interesting. So what advice would you have to self driving car companies, because I mean they are taking a little different approach, it seems like.

Mitchell Weiss: Yeah. Most of the self driving car companies are wrapping their success around LIDAR technology, which is better at ranging discreet points, but not as good as capturing lots of data quickly. But we think and some people are saying – as Elon Musk is saying it, a few others are saying it in the industry that they should start looking towards the camera technologies and we think that’s something they should be looking at as well and we think it’s something we might able to contribute. Other advice I have for the auto industry, I’m on record for, we spend a lot of time working on the safety systems around our vehicles. So we run around over 450,000 miles with our vehicles and we’ve never hit a person, never injured a person. And there are industry standards in our industry, the unmanned guided vehicle business, industrial unmanned guided vehicles that require us to do detailed risk analyses and risk mitigation plans and design the systems to international standards of electrical reliability. I think the auto industry better start working harder on that stuff. You know just putting these things out in the field and hoping that you don’t hurt people and learning from it is a dangerous way to do it.

David Kruse: Definitely. And are you guys doing anything – and maybe this is confidential in the self driving industry now?

Mitchell Weiss: We are continuing to develop our technology and improving it at the time. One of the advantages to how we do stuff being so computer intensive and being based on camera technology is as the technology improves in the market place because of ubiquity in cameras and phones and the power of processing, we are continuing to be able to do better precession, higher range, higher density data.

David Kruse: Interesting. Yeah, that was a pretty smart idea about 13 years ago you guys had.

Mitchell Weiss: Yeah. So well one of the things that Hans Moravec did, you can look him up, is when he starting working on this stuff in the 70s at Stanford AI, he went to look to see how long it would take – how fast computing power was improving, you know based on Moore’s Law of doubling everything, but he compared it to biological systems. And somewhere around 2000, he kind of came to the decision that the time was right for given dollars worth of computing you can compute this kind of stuff and we’ve been very adamant from a product development perspective to make sure that everything we do is not based on being clever or inventive, even though we invent and we do clever things, but that it’s based on the benefits we can get from improvements in computer power. That’s why we say we are more a software perspective company than a hardware company, because if something is really hard to do today, in five years it will be mainstream, so we might as well do the hard things.

David Kruse: That’s a good way to look at it. And I know we’re almost out of time here, but I had a couple more questions if that’s okay and one of them is about kind of a use case. So you know let’s say you’re in a warehouse and we’re going away from the technical to practical, but so somebody would load up a pallet truck and then off of it goes. Is that how you could work or…?

Mitchell Weiss: That’s one of the use cases. So there’s two big use cases we have. One is in the distribution and warehousing space. So most people don’t know it, because it’s so invisible to us. But if you go to your local grocery store, they are getting a truck load of pallets everyday and that truck load of pallets is coming from some local distribution center and there are tens of thousands of these in the country, where your groceries get stacked up on the pallets and sent to the store and then they get put on the shelves. Those distribution centers are a half million to a million square foot facilities all over the country. They are all those buildings that look like white corrugated steel or just white windowless buildings that you see along the high way tucked behind the trees. In those buildings are guys driving forklift trucks from the loading docks, putting pallets away on shelves or taking the pallets off the shelves and putting them back out at the loading dock and they are travelling maybe 20, 30 miles a day on their forklift trucks. Every minute they are driving that truck around its costing a lot of money, so our customers actually measure their savings in how many cents per pallet they save moving them around. And if you’re moving 10,000 pallets a day, we’re saving a lot of money. That’s how the grocery industry works. We do a lot of stuff. In manufacturing – so that’s a win we have in distribution. Any of our people move materials around big facilities, we move them without a driver; pretty simple use case. In the manufacturing space, when you are building a car or a plane or a washing machine, and this is public record, so we are in a washing machine factory where we have 52 of our vehicles running. So the local remained with a supervisor and they are all pulling carts of parts to the assembly line, where the building over 20,000 washing machines a day. So from a sample storage area, parts are continually being brought out to the line and delivered just in time to where the people need them to build the washing machines. So that’s 50 tuggers running around, carrying parts of goods, instead of 50 forklift drivers with people driving them. And oh, by the way, the number one cause of industrial accidents in the U.S. that result in death are forklift accidents.

David Kruse: Really? Wow!

Mitchell Weiss: Yeah, a person every three days in the U.S. probably dies because of it. So in that washing machine facility their motivation to moving to automatic vehicles was a safety motivation. Turned out it saves them a whole bunch of money, because there’s 50 trucks driving around on their own, but their objective was to reduce the risk of injury.

David Kruse: Sounds like a no-brainer. [Cross Talk] Go ahead.

Mitchell Weiss: No, no, those were the two biggies; taking parts to where people need them or putting parts away somewhere.

David Kruse: And do you publicly disclose the cost for the vision system, because that’s kind of a loaded question which is the problem, because there’s soo many variables, all right. Fair enough, I probably won’t…

Mitchell Weiss: So we don’t sell the vision system on its own, we just sell the vehicles and we sell them in such a way that if a customer has an application, say from car companies setting up the power production line, they’ll come to us, we’ll send in applications engineers to work with them on the material flow requirements they have; we’ll figure out how many vehicles they need and then we’ll quote them with price for how much that would cost.

David Kruse: Excellent.

Mitchell Weiss: That and a magical secret so we can make money out of it.

David Kruse: That’s right, that’s right. So the last question, I promise, is a – yeah, I’m curious you know where do you want to take the technology at Seegrid over the next five years or whatever time makes sense? How do you want to improve it and what’s necessary to make that happen?

Mitchell Weiss: So there’s kind of a two-step action to that. Robotics, in our case materials handling technology, it’s been around for a long time and its main mission is to reduce the mindless jobs that people have, so they can do more mindful jobs and we work very closely to people where they are delivering goods to and taking them away from people. So we’d have to enroll the safety systems and stuff. Okay, that’s step A. Step B, its turning out that with the growth of e-commerce and all these large fulfillment centers going in, there’s not enough labor in the country to fill those jobs. You can get people in all these little town – they have a big distribution center that’s leading a 100 mile radius to deliver your next day peanut butter sandwich from Amazon or something. So it’s going to require giving our machines the ability to do more and then go from A to B. You’re going to have to be able to pick up the goods themselves, select the goods, maybe grab a whole case of goods or do something. So a lot of the development we’re going to do in the future is around what we call manipulation, load handling, whatever its worth; robot arms on vehicles or just the forks picking up forks or whatever that’s doing. So that a big part of it is giving machines more smart and more autonomy, not just in terms of path following, but in terms of following a function. Then there are other things we can do with the perception technology and three dimensional space in terms of building that using or applying it to other industries. Yeah, we don’t talk about it a lot. So it’s really about leveraging this ability to – if you think about what people can do by looking around and walking around, then you go to the next step of what they can do. I’ll now let them start to interact with the environment, because they know where they are and they know what they have to do.

David Kruse: That makes sense and that would give you an entire process from picking up the goods to delivering to the line, yeah that would be pretty amazing. That’s not easy.

Mitchell Weiss: To stock even the store shelves.

David Kruse: Yeah, that’s not an easy problem, but it’s a big one.

Mitchell Weiss: And it’s not one that you get into the business and expect to solve in a year.

David Kruse: No.

Mitchell Weiss: Its one of the other interesting things about the whole robotics and automation space. You better have it when you are planning in your mind.

David Kruse: Well, that makes sense. All right, well that’s a good – I like your vision. It’s probably a good place to end this podcast, but Mitchell I definitely appreciate your time and your thoughts and you’ve got a large amount of experience and we feel honored that you could share it with us.

Mitchell Weiss: You’re easily honored then. But thanks for your time and it was a great fun telling the stories.

David Kruse: Definitely and yes, we definitely appreciate it and thanks everyone for listening to another episode of Flyover Labs. As always, I really appreciate it and we’ll see you next time. Thanks Mitchell, thanks everyone. Bye.

Mitchell Weiss: See you.