 Hello global supply chainers, my name is Ahmad Hamati and today we have our last Hangout event in SC1X. In this course, we covered the key concepts in supply chains. We started with different methods for demand forecasting and we continued with inventory management, the techniques for inventory management and then we ended with freight transportation. Today we are going to talk about less mind delivery, which is one of the most interesting topics in supply chain management. I have invited Dr. Matthias Winkumbach and he is a director of Megacity Lab here at MIT Center for Transportation and Logistics and he is going to talk about less mind delivery. Thank you Matthias for being here. Thanks a lot. The stage is yours. Okay. Thanks also for stating that last night delivery is the most interesting part of the supply chain. I strongly agree. That's also one of the most challenging ones, which is probably why it's so interesting for us. If you think about why is last mile delivery considered probably the most complex, the most hard to optimize part of supply chain and why is it also one of the most costly parts of the global supply chain. There are probably three major trends that you need to keep in mind. The first one is urbanization. We see a constant growth in the number but also the size of cities around the world. I think in 2014 the number of cities with more than 10 million people was about 28 and the UN actually projects that to rise above 40 by the year 2030. That basically means we will have more extremely large urban centers and obviously this growth in urban demand and the growth of urban conglomerates doesn't only happen in terms of size. So cities usually can't just grow infinitely large. Instead, the larger they get, the more dense they also get to accommodate more companies and more people. And that is one of the key challenges in last mile logistics that is the density of demand. You have a lot of people to serve, a lot of stops to visit if you think about the vehicle routing aspect of last mile logistics. But also you are competing with more and more people around a very scarce resource and that is transportation infrastructure. So while the cities grow in density and in size, the transportation infrastructure so literally the roads in the city don't usually grow at the same pace. So things are getting more congested than what we see every day when we go into these cities. We have congestion, we have pollution. We have relatively low travel speed, relatively low accessibility. And that obviously poses tremendous challenges to logistics companies because for them it's going to get more and more difficult to do efficient routing to be on time when delivering to customers and the like. At the same time, urbanization has another effect, namely so I think the McKinsey Global Institute a few years back projected that about 60% of the global GDP grows until 2025 is going to come from only 600 cities. So the 600 largest cities in the world are going to account for about two thirds of the global GDP growth and at the same time also host about a quarter of the global population. Now if you think about what it actually means to be more productive in your last collaborations that means you have an immediate economic impact but also immediate impact on the lives and the livability of cities for a majority of the population around the world. The second challenge that we see comes a little bit more from the e-commerce side. So since e-commerce volumes are growing more and more, that obviously means that also the shipments that go into it are getting more and more fragmented. So while a couple of years ago, logistics service providers were mostly concerned about getting relatively large consolidated shipments to relatively few nodes within separate cities. Now that we are seeing a growth in e-commerce, more and more home delivery taking place, they have to basically take care of smaller shipments, smaller packages going to more recipients which again makes the routing aspect of last month logistics more and more challenging. Yeah, actually as you said e-commerce is one of the new things that we have right now. I think 10% of the purchase comes from the e-commerce online services so that's a big challenge for them. Exactly and if you think about some people might underestimate the complexity of actually doing efficient routing in the last mile setting. So for instance UPS in the US has about 120 stops per route on an average route setting. Now that doesn't sound a lot but if you think about the possible combinations of visiting those 120 stops you end up with a huge number and obviously only very few of these combinations are feasible but still you have to figure out which one is the optimal one. And I think the Wall Street Journal made a simple calculation and figured out that if UPS only got a little bit more efficient in doing this routing so if they only reduced their average route length by a mile or so that would translate into annual savings for UPS of about 50 million dollars only in the US. So that gives you an idea of how little efficiency gains already can have a large economic impact in last month. And the last thing that I wanted to touch upon is actually customer expectations. So like you and me we are getting used or more and more used to having faster deliveries, having more precise time windows, so more reliable, more controllable delivery services. We are getting used to being able to customize the delivery late. So let's say I order something from Amazon today to my home and I figure out sometime in the morning well I'm not going to be home in the afternoon so why don't I just change the delivery appers to my office. We are slowly but surely getting used to this being possible and also being almost free of charge but the complexity and also the flexibility that is required from a logistic service provider to actually deliver such a service is tremendous and so while the customer expectations are rising, they're willing to pay, it's not necessarily rising as fast. So cost is becoming more and more an issue for a logistic service provider to be efficient in the way they do last month. And another issue could be the transparency and kind of tracking services. So that's one of the things. Exactly, which boils down to the whole aspect of data. How do we use data? How do we use new technology to first of all become more efficient in our own operations internally but also to provide more service to our customers by giving them more transparency over the delivery process. Because sometimes customers are probably not that concerned about getting a delivery like half an hour earlier as long as they know when they're going to get it. There's nothing worse than promising a five hour delivery window to a customer and you show up in the very last 10 minutes of that five hour window. That's like super frustrating for a customer. So you rather tell him, well, we're going to show up within these five hours but we're going to tell you about half an hour in advance that we're now about to show up at your door. It gives the customer more control or at least a stronger feeling of control over the last more delivered process. It's cheaper for you, the last part of the service provider. And it's relatively easy to implement nowadays given the mobile data technology and sensor technology that's out there by now. No, Matthias, our learners have different backgrounds and they have learned about the TSP problem in SC0, the first course in the micrometer pipeline. So they know what it means, but they only know how to solve it with five nodes. So can you tell us what's the difference when you have the larger scale problem and compared to the, what are the challenges for larger scale routing problems? Yeah, I mean, if we, like most of the projects that we work on with companies around the world on improving their last mile operations are at least looking at an entire city. So you're no longer talking about, I don't know, TSP problems with 10 customers or something, you're talking of thousands of customers every day. And as I mentioned before, the routing problem, so actually the TSP is kind of the simplified version of what we actually need to do, which is a vehicle routing problem, where you have capacity constraints on your vehicle and the like, mother of problems. But as I mentioned before, this is an extremely computationally complex problem. So if you are dealing with thousands of customers every day, that's just computationally almost impossible to do this on the scale of an entire city. So we have to come up with different methods to basically help our project partners to become better. So for instance, if we do network design problems, we usually formulate these as what we would call mixed integer linear program models. So basically mathematical optimization models that try to decide things like where should I locate my distribution center? Where should I have satellite facilities for transport on shipment, for instance? And which part of the city should I serve with which vehicle and ideally also which customers should I serve from which facility with which vehicle on which route? Well, the routing aspect can't be done on an entire city level. That's why we use approximation techniques. So we use a method that the learners might have heard about already, a continuum approximation. So we do not do the explicit routing. So we don't solve the individual routing problem for every vehicle explicitly. But we make certain assumptions about how customers are distributed in space. A very basic assumption would be they are somehow uniformly distributed in at least parts of the city. And we make other assumptions about the properties of these customers. Like what is their drop size? So how much do we actually have to deliver to each of them and the like? And then we can formulate a closed form approximation of what we think it will cost us to serve these customers on an optimal route. So we basically take the characteristics of the demand. We take the characteristics of the facility that we're starting from. We also take the characteristics of the vehicle that we use in speed and capacity. And simply put, we plug this into a formula that gives us a relatively close estimate of what it will cost us to serve all those customers assuming that we do optimal routing. Great. So sorry for interrupting you. Just to make it clear, as I just said, and we know that it's not possible to solve to optimality the larger scale problem with mixed integer programs. Yes, right? That's why we use this approximation methods and we lose the optimality, but we, you know, at least we have some estimation. And usually that's fine because when we do, for instance, distribution network design. So when we redesign the facility infrastructure that a company uses to serve a city, we actually don't care whether this facility that works perfectly on a particular day because this is a strategic decision. So we rather think about demand as an aggregate thing. So we look at basically how demand overall develops over time and we try to design networks that serve the demand really efficiently overall. So not on only one particular day, but basically we look for networks that perform as good as possible across the entire year, for instance. So that's why it would also make a lot of sense to solve the routing aspect of it for one particular day to optimality because the next day is going to be different. So you're more interested in things like average density of customer in a certain area or average drop sizes rather than the particular realizations of one particular day or your demand. Yes, so as I understood, so we have two different aspects here. So one is strategic level. Then for solving a very, very large scale, a very large city. So you try to approximate some routing costs. But when it comes to the operational level or delivery, the investment delivery, you need to find some kind of best routes. Yes, for operational decisions, obviously you at the end have to run a vehicle routing problem. But this one is much more constrained because once you know where your facility is and which service area these facilities actually serve, then you're no longer looking at thousands of customers, but you're probably looking at 100 or 200 at most. And then you solve that for a relatively limited number of vehicles. That still takes some time, but that's solvable to at least close to optimal. How about environmental issues? Environmental issues, yes. It's becoming more and more important. It depends a little bit on whom you ask, obviously. But for instance, we right now doing a couple of projects in Latin America where there's much more awareness by now than probably a few years ago about the necessity to not only optimize the operations for cost, but also for, for instance, CO2 emissions or particulate matter emissions. So we use scientific methods like, for instance, the NTM method to approximate the emissions that emerge from our vehicle routes. And then we can basically choose what to optimize for to what degree. So whether to do pure cost optimization, to do pure, let's say, emissions optimization or something in between. I think we've got some kind of application of this rather than, you know, kind of UPS and this postal service. Is there any other application for last month's anniversaries in other companies or any other type of, you know? Yeah, I mean, the UPS is obviously the most straightforward application. But we are actually working a lot with manufacturers. So for instance, in Latin America, but also here in the US, we are working with, for instance, Coca-Cola or as a bush. So two large beverage companies. And their operations obviously differ quite significantly from the one of, let's say, a typical postal or express service provider. Because they deal with very heavy stuff. So they are bound to have vehicles that can actually carry the load of several cases of beer. It's a very manual process. So we really have to take care that we also parameterize our models. So there's a lot of data collection involved to really understand how much time do they need to spend at each customer to deliver, let's say, four cases of beer to a certain readily retail outlet on the street. And so we really have to also understand what drives route efficiency. So what customer characteristics, for instance, define how much time our delivery crew needs to spend to deliver to a certain customer. Or how does the road network topology actually affect the efficiency of our routes? Because as the bigger your vehicle gets, the more constrained you are in terms of how you can maneuver through a road network. That, for instance, affects one-way streets that these affect different vehicle types differently. If you have bad road quality, which we often face, for instance, in Latin America or Asia, then some vehicles are just not made to go on unpaid roads, for instance. And all of these things need to be accounted for, again, in our models. So we need to inform our optimization models about the available infrastructure, about the topology of the road network, about the properties of our customers so that we don't just model something, but we actually model reality. And that's the challenge, and that's where we strive for using the available data that's out there from various sources as effectively as possible to inform our models better. So you always talk about the trucks, about the other modes of transportation. You want me to talk about drones, right? Yeah, recently we have heard about these drones. So do you consider this as a part of the last mine delivery? Well, I hope nobody from Amazon is watching, but actually we have our doubts, let's say. So drones do have some applications, but we do not see them as a mass deployable solution for dense urban centers. And that's for various... As of the regulations or other issues? I mean, so first of all, technology-wise, it's possible. Yes, that's clear for now. But regulation is probably one of the largest hurdles that would need to be overcome. But with regulation comes immediately one other aspect of the platform, which is control. So for instance, right now it's probably easy to control, let's say, the delivery airspace if there's only five drones buzzing around. But when I talked about e-commerce volumes going into urban centers earlier, the volumes that we see going in there would require thousands of drones to go into those cities. And then you need some sort of control. And either you try to tackle that through standardization. So we would have to make vendors agree on common ways to communicate or to let the drones communicate with each other, which is challenging. The other security... Yeah, cyber security is on other things. Cyber security, yeah. Or just not even cyber security, but let's say we're talking about delivery of beer and you have a drone and it drops a case of beer on pedestrians on the street. That is very unpleasant. So there's a lot of safety and security risks but probably the lightest hurdle would be total acceptance. Because right now everyone thinks, oh, it's cool that Amazon does those drone deliveries. But imagine you have thousands of drones buzzing around your head. I don't think that people would really want. And it's also not the most efficient way of delivering. What we do see, for instance, Mercedes-Benz has released a couple of months ago a concept of an electric truck. I think it was even autonomous truck. It would kind of serve as a mothership for a drone. So that truck would actually have inventory in its back. It would have two or three drones on the top. And the truck would basically circulate through the city depending on where the amount is expected. And then the drone basically can just take off from the truck to do the very final delivery, like a few hundred meters to the customer. That is a solution. I mean, that's still probably 10, 15 years down the road. But that's something that could make sense because you minimize the distance that you have to overcome with the drone. If you think about energy efficiency, lifting stuff in the air and keeping it there is the most energy-intensive thing you could do. So you want to minimize the amount of distance that you have to overcome that way. What we do see much more application of unmanned aerial vehicles is actually not in cities, but in rural areas. For instance, DHL has been piloting their drone deliveries to, for instance, islands of the coast that are very sparsely populated. So it doesn't really make economic sense to go there. But especially as a post operator, you have to go there. You have to provide a postal service, or for medical supplies, you have to go there. You have to supply these people with medical services. And drones could be a much more efficient way of doing so than going there with a ship. Or in the mountains, you can actually go to remote places in the mountains more safely and more efficiently and also more quickly. With a drone, then if you had to go with a ground-based vehicle. So these are applications where you do see potential for that technology, but it also makes sense to do it in an urban environment, not so much. And maybe in some kind of elsewhere application. Yes, so, you know, universal. Yes. And that also usually goes along with higher margins on the merchandise that is being delivered. And then it's also easier to recover the costs that need to go into establishing an infrastructure for chrome building for them. Yeah, and also kind of emergency services like ambulance and these things. So it would be kind of another application of this let's call it not less my delivery, but the kind of route thing and this type of problem. Yes. In a way, again, you have to, if you design, let's say, a network of stations for emergency services. The problem is mathematically speaking, relatively similar to a scripted delivery service. You have to find locations that cover the entire demand. So everyone needs to be covered by emergency services. And you want to minimize the response time. So that's kind of a trade-off. And you probably have a budget constraint that you can't exceed. But basically that's the same problem that logistics providers face when they try to design networks for the delivery of goods. You start to talk with the demands part. So demands is kind of input given data to the problem that you are dealing with. And we have some kind of, we start our course with demand for this. So, and what kind of problem or kind of what kind of issue and challenge are regarding the right input for your algorithms and your problem that you are dealing with. Well, I mean, one common issue that you always face when you start a project is getting the right data at the right quality. So we usually work with a lot of complexional data from our project partners. And that's literally just in a download from their order management systems. So we look at order data and try to understand what did they deliver in the past to whom and when. And then usually you have noise in that data that needs to be cleaned. You have to geo reference that data, which they usually don't do because they don't need it. But when we model distribution network design problem for a city, we basically need to locate every demand information with an X and Y coordinate in that city. So that's a challenge that we are typically facing. And then obviously it's relatively easy to look at historic demand patterns. So if we get multiple years of data from our project partners, we can very nicely characterize their existing demand seasonalities, for instance, or the existing distribution of demand within space. So there's a spatial and a temporal component to how demand evolves over time. It gets more challenging if you're trying to extrapolate that into the future. Because right now we are talking to a company in Latin America that provides paper products like diapers and the like. And they've been traditionally serving large retail stores. But in Latin America, the traditional retail sector is extremely dominant. So we have a lot of very small, usually single owner operated retail outlets. And they think about entering that market. So serving those traditional retail outlets directly. So they have no historic information about what the demand from these places might look like. And that's where we try to help them by coming up with proxies. So for instance, by looking at the distribution of population density within cities or the distribution of socio-economic data, like income level household sizes, to come up with reasonable estimates for how the demand that they need to serve might look like if they enter in this new customer segment. That's the challenge, especially because this socio-economic data is hard to get at the level of resolution that we're trying to obtain. So sometimes we even have to send people on the ground to do some kind of survey studies to get a better understanding of how the retail landscape but also the consumer landscape in different parts of a city actually look like. Great. So I think we are running out of time, but is there any last words that you want to say? No, I can only encourage you to study last my logistics even more because as you're rightfully started the chat, this is probably the most challenging for lots of people, the most interesting part of a supply chain. And it's great because it combines, I think, three things that I at least find interesting. One is dealing with large amounts of data and different sources of data, dealing with mathematical modeling, trying to basically give people tools that help them make better logistics decisions. And at the same time, even though that's all mathematical, it's very tangible. So we deal with moving stuff around. We deal with serving physical customers. We deal with services that we experience on a daily basis. And that's what I like about this field and I encourage you to study it more. The cost-saving part is the most important for the company. Thank you, Matthias, for being here. This is a great experience. Thank you for sharing your experiences. So we have a short break and then please use the breakout rooms just to give your feedback about this talk and the general feedback for the course. Which part of the course you like the most and which part you think that it needs improvement. So we will be back in maybe 10 minutes. Okay, thank you.