Speaker 0 00:00:00 Ambi robotics is about empowering people to handle more with AI powered robots that pick and place.
Speaker 1 00:00:13 Hello everyone, and welcome to the robot industry podcast. We're glad you're here and thank you for subscribing. My name is Jim Beretta and I'm your host. I'm super excited to have Jeff Mahler
Speaker 2 00:00:25 On the podcast today. Jeff Moore is the CTO and co-founder of Ambi robotics. He received his PhD in AI and robotics from the university of California, Berkeley for his work with professor Ken Goldberg on the dexterity network, a state-of-the-art method for rapidly teaching robots to pick up items that they've never been trained on before his robot vision research has appeared in the New York times and MIT tech review and Forbes, and has been nominated for numerous awards at international research conferences, including best manipulation paper and best human robot interaction paper. Jeff also holds a BS in electrical engineering from the university of Texas at Austin. While there he was the co-founder and lead computer vision engineer at Lynx labs, a 3d scanning startup that was acquired by occipital in 2015. Thanks Jim. It's a pleasure to be here. So
Speaker 1 00:01:24 What do you do now with Ambi robotics?
Speaker 0 00:01:27 MB robotics is about empowering people to handle more with AI powered robots that can pick and place, uh, virtually any item. Um, and, and we can do this by actually configuring robots out of building blocks of various hardware components like robot, arms, grippers, and cameras to implement and transform logistics processes so that robots can work together with people, uh, to multiply productivity for processes that today are becoming very hard to sustain with the growing volumes in e-commerce. And in fact, many of these processes, um, are, are finding a very hard time attracting and retaining manual labor, um, primarily due to the tediousness of the work, as well as the very high injury rates. Um, some of these jobs are actually posting as high as 10% injuries across their workforce a year. And as a result, there's very high turnover. So people are leaving these jobs and droves usually within the first two to four weeks.
Speaker 0 00:02:24 And many of the people we've talked to have had open positions for a very long time. Um, and for some of those positions, they've had to hire five to 10 people, uh, for every position that they're going to keep one full-time person in such tremendous amounts of turnover. We're trying to, oh, go ahead. No, that's crazy. It is. And it was actually shocking to me to go behind the scenes and see that, uh, you know, e-commerce is sort of a black box, I think, to the average consumer. And, and we don't realize that behind the scenes there's, there's so much going on so many touches of individual items as they go from our virtual shopping cart to our front door.
Speaker 1 00:03:03 Yeah, it's something certainly we don't think about, about the, the carpal tunnel and just the frustration out there. So how did you get started?
Speaker 0 00:03:12 So, uh, my, my start actually came, uh, when I was, um, working on the 3d scanning company, Lynx labs, uh, back at, uh, university of Texas, um, and sort of seeing that there was an opportunity to really bring a 3d technology and commercialize it. And, uh, we were at that time using new scanners, like the Microsoft connect to scan interior spaces for things like architectural redesigns. Um, we ultimately sold that company, Chuck <inaudible> and I went on to UC Berkeley to study robotics, seeing a big opportunity to connect those vision systems with things that can move and interact with the world. And while there, um, I joined professor Ken Goldberg's lab and came across this fascinating problem, uh, called universal picking, which is how to enable robots to rapidly and robustly handle a wide variety of items that they've never encountered before in the real world, pick them up and place them in any location.
Speaker 0 00:04:07 And what I found so captivating about this problem is that, uh, it's something that we often don't actually think of as, as being a sort of landmark of, of human capability or intelligence. We often grab sings, manipulate them without even really thinking about it too hard. But actually when we try to get robots to do these sorts of tasks, it's tremendously difficult. Um, we're dealing with a lot of limitations in sensory capabilities. Uh, what sort of sensors are returning in terms of noisy data, as well as lack thereof of data, you know, with occlusions as well as the inability to control, um, and actually robustly sort of reacts to the world. Uh, so robots don't have hands, but we're often using different manipulators, like a two finger gripper or suction cups, which are highly reliable, but also, uh, don't have the 26 degrees of freedom our hands do.
Speaker 0 00:05:01 Um, and on top of that, most robots are effectively numb. We don't, we don't really have robots skin anywhere near the level that, uh, humans are capable of, of reacting with. And so all of these factors compound into making robotic, grasping and manipulation a tremendously difficult problem. So I started working on a methodology for trying to enable robots to take one step further in terms of their capabilities. Uh, traditionally at that time, robots could pick and place items, um, in industrial automation when they had precise knowledge of the items that are going to handle a CAD model, uh, perhaps the mass and material properties of those items. Um, and they could repeatedly sort of execute the same motions on those items over and over again. What we wanted to do was enable them to handle new items that they never encountered before, without any prior knowledge of those items in advance.
Speaker 0 00:05:57 And we saw a big opportunity to take the advances that were happening in deep neural networks at the time from computer vision, um, and apply them in robotics. All these great results were happening in the early 2010s from classifying images, uh, detecting items in images and even playing games like Atari and the game of go at super, super human levels. What we ended up finding though is that this is very difficult to apply to robotics largely due to the challenges of collecting data for the robots. And so we started studying how to enable the robot data collection, collecting tens of millions of examples, of grasps on different items, uh, to be more automated and to be much faster and through the decks net project or the dexterity network. Uh, we came up with a way to do this actually by entirely training a robot to grasp items and simulation analyzing tens of thousands of 3d CAD models, uh, various contact points on the items and scoring them with analytic grasp quality metrics that could say whether or not it grasped might succeed or fail.
Speaker 0 00:06:59 And then we could tie this into rendered images, um, that has grown as a technology to largely video games to generate these large data sets of images and grasp pairs that could then be used to train vision systems that ran and handled new items on day one in the real world, we had a lot of success through that project, um, and, and had a series of papers showing how this could be applied to many different, uh, sort of grippers and even on several different types of robots. And through that, we received a lot of press and even got invited to present the system to Jeff Bezos at his Mars 2018 conference. So at that time, uh, we were realizing that there was a huge opportunity to go beyond academia and take this technology into the real world where it could make an impact, um, on various problems and where we saw the biggest opportunity to make an impact was in logistics. Whereas I described earlier, there's this huge pain point of attracting and retaining labor to do these jobs and a big opportunity to provide robots, to work with humans, to multiply productivity in the warehouse. So we left campus in 2019 and we've been building a business since then, and it's been very exciting to see all the progress in the industry. So
Speaker 1 00:08:10 I guess your ideal customer is the logistics markets and the logistics operators.
Speaker 0 00:08:17 So today, yes, we are serving in logistics, doing tasks like parcel sortation and kidding. Uh, but more broadly speaking, uh, we are on a mission to help any company that is having these challenges of attracting and retaining labor for these pick and place tasks. Um, and we are working with, with companies today and are excited to work with new customers that are really thinking about how to transform these problems now and revolutionized their processes, uh, going into the future. Uh, so we're, we're looking at extending applications beyond parcel sorting and kidding, uh, but we have a start there and we're really excited about the huge opportunity within those markets alone,
Speaker 1 00:08:55 Problem beyond, uh, beyond labor attracting, retaining, uh, talent, uh, is there, is it speed as well? Of course.
Speaker 0 00:09:03 Sure. So, so this problem sort of ties into capacity. So how many parcels or, or orders that can be fulfilled by April filament center in any given unit of time, like for example, how many orders can be fulfilled in a day, um, and that's, uh, can be addressed by, by increasing speed. Um, and, and that is where the robots can help by actually, um, changing the way these processes are done. Um, the effective throughput of every person working in the warehouse ends up going up and so we can increase capacity significantly. Another area actually is an accuracy. Um, some processes that are done very manually today end up having accuracy issues, uh, packages get sorted to the wrong destination, um, at surprisingly high rates, uh, sometimes, and being able to ensure that packages go to the right destination, um, give people a better customer experience and can also be critical in, in applications where there's a real need to get, um, items like food and medicine to the right location on time.
Speaker 1 00:10:06 And so are you robot agnostic at Ambi?
Speaker 0 00:10:09 We are our software, uh, can be configured from a library of different hardware components, such as different robot arms and have our own tools like suction cup or two finger grippers, um, and various different depths and stereo cameras, as well as even traditional automation equipment. We interface with, uh, conveyance, um, as well as, um, material flow mechanisms for say bringing bins to and from the robot and this software, uh, for tying these components together and making them work seamlessly is called <inaudible>. That is our software operating system for configuring systems from these various components, um, and ensuring that they can be quickly running on day one in the warehouse.
Speaker 1 00:10:50 And one of the things we talked about, uh, before the pod started, it was adaptable AI, what is adaptable AI
Speaker 0 00:10:58 Adaptability, I is, is about exactly this concept. Being able to take an AI system and have it, uh, exist on a specific hardware configuration and work very rapidly. Uh, traditionally sort of computer vision AI systems are trained on a particular task from a large, and in many cases, often static data sets. Uh, but what we're dealing with in robotics is providing robot that might be picking and sorting parcels out of a bin. And then we may need to create another robot that actually picks those parcels off a conveyor, um, and sorts them into a tote rather than into a mail sack. And being able to make these changes rapidly is really what adaptable AI is about and the way we achieve that is I a combination of simulation reality technology, um, stemming from the deck snap project work, as well as modular hardware, this library of hardware components that we support and can be rapidly designed, uh, to implement these different processes.
Speaker 1 00:12:00 And do you, so do you use a simulation software or in which one would that be?
Speaker 0 00:12:06 We do leverage simulation software quite heavily. So our simulation environment, uh, is crafted from both, uh, open source components as well as our own proprietary simulation tools. Um, and on that side, a lot of our developments are particularly around modeling contact. Um, how robot fingers, if you will, like suction cups, uh, form a seal with items and actually apply forces and torques to them. And the second main component is domain randomization, how to actually randomize the various items and their material properties and appearance, as well as the physical properties of, of items in the world, such as friction in order to ensure that what is learned in simulation transfers robustly in reality to reality, excuse me. And, and the reason this is important is because simulation, uh, is never modeling the real world perfectly, but what we can do is actually make the simulated world effectively harder than the real world is by randomizing things, um, in such a unique way and challenging way for the robot, such that when it's encountering items in the real world, um, it's essentially sort of a subset of what it's seen in the simulated world. Um, in terms of the open source components we build on, on tools like bullets for dynamics simulation, um, as well as rendering tools like open GL, uh, to help us build a complete system that can rapidly, uh, simulate and scale up, uh, to train these models effectively, overnight in about eight hours. Nice.
Speaker 1 00:13:39 So just to switch the conversation a little bit, how important is data for your customers? I mean, data is obviously important for you, but what does that conversation sound like with your customers?
Speaker 0 00:13:52 It's usually important for our customers. And in fact, one of the things we've discovered when talking to folks in the market is that often there are data opportunities that, um, may not have even been realized by the customer. For example, they are, uh, tracking all of their own processes today, measuring rates and how many items are moved out of the door in a certain period of time, but they often lack a finer level of granularity. Um, in terms of, you know, precisely when was an item picked and how long was, say the line blocked at this particular point in time. Um, and one of the exciting things about robotics is that we're naturally having to gather that data as the robot interacts with people to implement a particular process. So we have a large amount of timing, data to help identify bottlenecks and productivity enhancements for the customer, which they really love.
Speaker 0 00:14:45 Um, and a second benefit is collecting a lot more data about the items that move through a customer systems. Knowing what items are being handled is very important. And many companies do of course have existing infrastructure for these problems. However, we can give them a finer level of granularity, um, imaging items as they're in the robot hand. Um, dimensionalizing those items, um, measuring the weights of those items. We can verify that back to you say, uh, what's being charged in terms of weight for, for packages and shipping, um, and even give customers better visibility on what types of materials come through their system are their parcels, uh, coming through the stream bags, or are they boxes or envelopes, we can help, uh, reveal the answers to some of those questions and help them optimize their processes into the future to better handle those types of items. So if
Speaker 1 00:15:37 They see that their packages are getting bigger or smaller or that, so they can take action.
Speaker 0 00:15:42 Yes, exactly. It's empowering them to take more action based on data that they never had before. And any
Speaker 1 00:15:49 Other problems that you're solving than just picking and placing.
Speaker 0 00:15:54 Yes. So, uh, we really think of ourselves as, as a service company, we, we implement services for the customer and that involves robots, um, doing various types of tasks. So we pick and place items, but we really think ourselves as also sorting items, uh, kitting items, you know, orienting them, uh, to fit within a constraint location, and even packing items. Um, and so it, it really does go go beyond just the pick and place. Um, also of course, you know, going back to the data question, providing insights and data that the customer never had before, but our sort of fundamental technology are built on starts with the ability to pick and place these new types of items that the robots never been trained on before. Very cool. And can
Speaker 1 00:16:39 You share any customer success stories? I saw one recently on your website.
Speaker 0 00:16:44 Yes. I'm really excited to share that with you. Uh, we recently announced our partnership with Pitney Bowes, uh, to implement parcel sorting robots throughout their network. And this is really exciting because, uh, Pitney Bowes, uh, while historically largely serving in sort of, uh, postage metering has been transitioning over the last decade into a third-party logistics company. And their business has grown like crazy, especially over the past five years. And we're helping them feel that growth and demand, uh, by providing robots that multiply at the productivity of operators in their warehouse. So we installed our first robot, um, in their facility just about a year ago. It operated through the peak season, uh, 20, 28 had hundreds of thousands of lives, sorts to them. And now we're scaling up production over the next several years to actually scale these robots throughout their network. And we're already seeing major gains, uh, over 50% higher throughput and a three times higher accuracy rate with the robotic system.
Speaker 0 00:17:45 And we can continue to prove from there. One of the things that's excited is seeing exciting is seeing how people use these robots and, and how the operators on the floor are really excited to interact with the robots, um, and actually sort of transition away from moving these packages around with their upper bodies that can weight over 20 pounds and can cause injuries to actually sort of ergonomically handling bags that come off of a cart, um, and placing those into their destination to go under a truck. Uh, so it's been very exciting to, to really, um, actually grow and, and, um, build that human element, that human interface as well together with Vinny voce.
Speaker 1 00:18:27 No, that's very exciting project. I was going to ask, how do you work with your clients? So you, you obviously in this case worked directly with Pitney Bowes, do you, but do you work with integrators or other companies?
Speaker 0 00:18:39 Yes, we work with a network of different partners. Um, we have, um, build partners that we work with. We actually recently announced our partnership with DW Fritz who is helping us build and scale these robots. Um, and they do a lot of systems integration in a more sort of traditional industrial automation and are looking to actually expand their business into the logistics world where, um, these new types of robotic automation are really starting to grow and expand. Uh, we also work with a number of partners to help us install and service these robots. Uh, so we're certainly not operating entirely by ourselves and trying to leverage the network of, of excellent, uh, partnerships that exist out there today. And I see
Speaker 1 00:19:22 That you're hiring a lot of software engineers. I thought it was sleuthing your website.
Speaker 0 00:19:28 Yes, we are. Uh, we are really excited to be growing and we need more people to join the team to, to help us build these robots and scale them. Uh, we have a lot of exciting problems to work on, on both the robotics side and in new areas, such as lead management tools to help manage all of these data points, such as images and videos, uh, coming off the robots. Uh, so we're hiring in a variety of roles, not just in robotics, but also those with expertise in things like backend engineering, uh, user interface, data management, and security. Um, and we, uh, are a team that's, that's really, um, excited and passionate about what we work on. Um, and we're, we're looking to find other individuals that share that sense of passionate about robotics and making an impact, uh, want to have a high level of ownership of what they do. And also have a great sense of adventure in, in coming into this new and exciting industry and, and helping us, uh, really shape that as we continue to grow. Yes,
Speaker 1 00:20:26 It's a very exciting industry. So Jeff, thanks for coming onto the podcast. I was going to ask you, what do you like to do when you're not in front of customers or in front of robots?
Speaker 0 00:20:37 Great question, Jim. Um, I'm a big fan of the outdoors. I head out to the mountains often to do things like trail running and mountaineering and skiing. Um, I also am really passionate about music. I play guitar and drums and like to, um, uh, improvise with, with some friends to, uh, create musical pieces. Very nice.
Speaker 1 00:20:57 Jeff, how can people get in touch with you if they wanted to follow up?
Speaker 0 00:21:02 Uh, you can email me
[email protected] or connect with me on LinkedIn, um, would love to hear from you and, um, start a conversation.
Speaker 1 00:21:13 Our sponsor for this episode is Earhart automation systems, Earhart builds and commissions turnkey solutions for their worldwide clients. With over 80 years of precision manufacturing, they understand the complex world of robotics, automated manufacturing, and project management, delivering world-class custom automation on time and on budget and contact when their sales engineers to see what Earhart can build for you. I'd also like to thank our partner, a three, the association for advancing automation, 83 is the leading automation trade association in the world for robotics, vision and imaging motion control and motors, and the artificial intelligence technologies visit automate.org toward more. I'd also like to recognize our partner painted robot painted robot builds and integrates digital solutions. They're a web development firm that offers SEO and digital social marketing and can set up and connect to CRM and other ERP tools to unify marketing sales and operations in
[email protected]. If you'd like to get in touch with us and by us, I mean me at the robot industry podcast, you can find me Jim Beretta on LinkedIn. We'll see you next time. Thanks for listening. Be safe out there. Today's podcast was produced by customer attraction, industrial marketing, and to thank my nephew, Chris gray for the music, Chris Colvin for audio production, my partner, Janet eighty-three and painted robot. Our sponsor Earhart automation systems.
