Matthias Winkenbach - Last-mile logistics on steroids

super well okay everyone welcome this is the MIT Mobility initiative weekly Forum our regular host Professor jinwa Zao is traveling today so I will be filling in for the role of host I'm John movinzade with the MIT Mobility initiative and my colleague move on at Laurie will be handling the Q a the curation of your questions that you enter into the chat function we have a very special guest today but before I introduce him I just want to reinforce the norms for this forum that jinwa typically talks about at the beginning which of course is camera on we want to see you voice off and the key word to remember with this form is engagement so we ask you to please enter your questions and your comments into the chat we do share those chat feeds with the with our weekly Forum participants and they do find the comments really very helpful useful interesting so so please type in your your thoughts uh as we as we go along uh now just to give Matthias a sense of who's actually with us today I see we're up to about 185 participants some of you have already done this if you could just quickly type into the chat your organization where your which organization you're with and your current location so let's see I see um course MIT in Cambridge Berlin Seattle New Hampshire Argentina Poland London Vancouver Canada Montreal Cambridge Bogota Chicago Atlanta Tokyo Japan Burns Switzerland Knoxville Tennessee Fargo North Dakota fantastic okay well that is a quite uh diverse group of participants as always so as you can see Matthias were straddling uh multiple multiple time zones here so um I will do my best to uh introduce uh Matthias winkenbach who is the director of the MIT Mega City Logistics lab and a principal research scientist at the MIT Center for transportation and Logistics his current research focuses on multi-tiered distribution Network design in the context of urban Logistics and last mile delivery Urban Freight policy and infrastructure design as well as data analytics and visualization in urban Logistics context Dr winkenbach received his PhD in logistics and His Master's in business with specialization in finance and economics at whu the auto bison School of Management in Germany he also studied at the NYU School School of Business the stern School of Business in New York as well as Usher say in Montreal Canada his doctoral studies focused on the optimal design of multi-tier urban delivery networks with mixed his work was closely linked to a research project with the French national postal operator friend to the to the MIT Mobility initiative and we've had the pleasure of working together on several projects Matthias it's a pleasure to welcome you to The Forum the floor is yours thank you John um and I hope that my connection is isn't the problem here because we were breaking a little bit in between and I I hope it wasn't on my end but either way welcome everyone thanks for having me um in the in this kind of circle and I guess today before I start this is not going to be a very technical talk I basically wanted to give a bit of a broader perspective on how we currently see last my Logistics evolving and it's evolving relatively rapidly and I wanted to kind of pinpoint one or two topics that we believe are particularly relevant here and where my lab is also actively doing researching and certainly quickly show my slides and then we can take it from here and please feel free to ask questions in the chat and then I guess we'll probably moderate that and if there's a recurring theme feel free to also interview me during the presentation but yeah uh General asked me to come up with a very catchy title and that was a few months ago so I called it last mile on steroids there's nothing about steroids in this talk um but basically what I wanted to somehow kind of get across in this talk is that we have to connect multiple things to actually design the future kind of lost my Logistics systems that we will all be experiencing as con inspiring experiencing as consumers for instance and that is collecting human knowledge human experience with What machines can nowadays do and basically Advanced algorithms AI that ideally try to connect the other two but before I go into the details of this presentation um let me start with a few sorry I have a bit of a quote with a few high level thoughts um one is and these Styles will basically guide you through this entire presentation but first thought is Last Mile Logistics is still probably the key value driver for especially the e-commerce industry and most importantly is a key value driver in the kind of on-demand economy that we're in where everything needs to be available basically our talk when people order it whenever the people people feel the need for certain products it needs to be available so they're kind of on-demand mentality is only possible to address for the e-commerce industry through appropriate last night Logistics systems the second thought throughout this presentation is that I mean obviously all of you have probably played around with Chef gbt lately but I think we're at a point where we can safely say that apart from all the very good and important work that has been happening over the last couple of decades in the on the or side of things so when it comes to optimization methods and the like probably the biggest Frontier of disruptive lost my Innovation and therefore probably also the biggest Frontier for kind of disruptive future research even lost my Logistics is AI and what personally I would like to see though is that we don't just see AI as something that will eventually replace human capabilities in lost my Logistics systems but rather as something that can augment what humans basically can do and what humans also do physically within last night Logistics operations um the third point I mean we love to talk about drones we have been talking about drones for many years in a way drones is just a placeholder for all sorts of autonomous Technologies to be delivery robots be it kind of uh assistive technology for humans that are doing some sort of physical task in the last minute Logistics system but for this talk let's talk about drones and how they can actually boost Last Mile productivity and flexibility but also what they cannot do and what they will likely never be and for the example of drones for instance here the high level thought is we don't see them as a standalone customer facing solution in lost my Logistics we do see them again as something that could complement existing networks could complement existing delivery models and I'll show you a few examples for that in a bit and then taking basically all of this together taking a step back and thinking about okay where does The Last Mile Logistics industry actually need to go from an r d perspective we believe that um Innovation needs to be addressed a little bit more holistically than it is being done right now so a lot of love my Logistics r d is still going very much into one single Direction like let's find the next best routing algorithm or let's find the next best vehicle technology that we want to use in optimal distribution but basically what we want to bring these different perspectives together and most importantly we believe that it's not always just an either or decision for instance it's not just AI versus human decision making it's not just drones versus the good old ground delivery vehicle it's usually a combination of these Technologies the combination of these methods that can yield the best results for lost analogistics design but first of all um why do we um in here and I guess I found this really interesting quote um in a recent report by McKinsey saying that retail has experienced more change over the past five years than in the prior 50 and that is probably true I would just add retail Logistics has experienced more change over the past five years than in the prior 50. so um there's a lot of very Dynamic developments going on in the retail industry in particular predominantly driven by e-commerce obviously and um there's a lot of pressure to create Superior value not just to Consumers but also to businesses and society as we basically enter this next stage of online Commerce retail and only Channel distribution and that's why we need to basically fundamentally rethink about how we design lost my Logistics systems and also how we best leverage again humans technology and algorithms to make that work um some of you probably have not seen this before to some of you have even talked about this before I just want to very briefly wrap up like what are the main pressures that the logistics industry is currently facing uh kind of in light of a trend towards faster more flexible e-commerce uh distribution um and one big trend is obviously driven by us turned by the consumers consumers today typically demand instant gratification and it's pretty hard for delivery service to actually keep up with that like while a couple of years ago we all were happy with receiving things I know two three four days after having placed an order online that mentality is changing more and more people are just getting used to the fact that if I order something online today I'm gonna get it tomorrow if not today and there's also economic reasons why many retailers want to push um for the ability to deliver more frequently to deliver faster and to deliver through more than just the traditional physical retail Channel but basically to enter the online Channel space because there's several studies that show that only Channel consumers a purchase more frequently and they buy more they buy a higher value products and so a vast majority of retailers is actually looking into investing into Superior delivery delivery capabilities particularly to their online clients right now we're talking mostly about companies investing into the next day delivery capabilities but I'm relatively certain that in a couple of years from now we will mostly be talking about same day and potentially something day delivery services the challenge with that is if you just think about it think of the us as a market and if you wanted to serve the us as a market with something like a three-day delivery time you could basically reach about 80 percent of the US population within three days out of just three big distribution centers scattered across the country so very centralized distribution if you move to delivery speed of next day delivery and you still want to reach 80 of the U.S population you would already need more than 90 fees to achieve that and now if we take that one step program think about things like same-day delivery or sub same day or even instant delivery it's clear that this can no longer be sustained by a centralized distribution system as we know it but more and more companies are basically now investing into decentralized networks so multi-tier distribution networks where you probably still have big stock holding facilities somewhere in the middle of nowhere but we're actually the most most of the Fulfillment process is happening quite close to the consumer already so you have a highly fragmented distribution Network they've also highly fragmented inventories in a much more complex system and to manage and to optimize and that's where from a logistics research point of view also a lot of existing methods currently start failing because they were just not designed for this level of complexity and fragmentation um my slides got stuck there we go the next thing on this list is actually kind of related to also delivery speed but also delivery consistency or reliability because there's more and more evidence for the fact that customers no longer want things fast no not only one thing's faster and more flexibly they also want more certainty around that process they want to have more certainty around when exactly they're going to get their delivery they want more kind of surgery around whether that delivery is going to be successful or not whether the driver that's doing the delivery for instance is the same one that used to serve that customer the day before um so this is a lot about how can we make last night delivery services not just faster but also more reliable and the problem here is that as we move away from let's say a proprietary distribution by the big e-commerce retailers out there as we move closer towards what we would call the gig economy so we see more and more crowdsourced delivery services for instance one big element basically gets missing that would usually explore reliability and consistency and that is driver familiarity so there's many studies that show that if a driver is familiar with a certain Road territory that driver can deliver to those customers more efficiently more reliably and also more safely now if we move to a Gigan economy where basically it's every time it's a different person delivering it to a certain customer to a certain area that familiarity gets lost and here is where again algorithms need to be developed to kind of Leverage data to basically compensate for the lack of driver familiarity with a certain delivery Zone surprisingly all of this puts a lot of pressure on cost um of distribution services but also on productivity so to give you an example we've been working with large food delivery networks in India over the last couple of years and one of their stories was obviously how do we make food delivery more cost efficient but the even bigger struggle was how do we find enough drivers or delivery people to actually keep up with the rapid growth in demand and that's honestly through pretty much across the entire e-commerce industry where there's just a consistent lack of skilled Workforce so a lot of work has also to go into the question how do we leverage the existing assets that we have from existing vehicle of the existing driver crowd more effectively how do we make these people more productive while not jeopardizing on things like safety and reliability and last but not least sustainability is no longer just a password but it's actually a credible business objective of many of the companies that we work with while a couple of years ago there was always some statement about sustainability in the research segments that we basically created before starting a research project it was often just a lip service it was often not a lot of real commitment behind kind of a sustainability objective in the work that we did together with industry and then has changed dramatically over the last couple of years I mean we've all seen the climate pledge initiative where hundreds of companies have by now pledged to become carbon neutral by 2040 basically representing a huge traction of the U.S economy too and um so there is a credible commitment to becoming more environmentally friendly um the big challenges here from Logistics industry point of view is obviously technological uncertainty the sheer cost of transitioning very Capital intensive assets strong traditional Technologies to our less carbon intensive Technologies and then obviously also especially right now a rather unclear economic Outlook so here a lot of research has to go into how do we basically alleviate some of that uncertainty how can we provide companies with the right models to basically work around the uncertainty that they see around technology around cost around economic potential of these Technologies so that's just a level set like um like why why are we even doing research in this space and why are there so many interesting problems to solve um in this talk since we have just a little bit of time I want to dive into two very popular topics one is AI and machine learning and how they could help us in lost my Logistics and the other one is as I said drones or robots a lot of technology in general how could that type of Technology help us build the future of my Logistics systems let's dive into AI or machine learning person as I said before we we strongly believe that AI at least in our industry isn't there to fully make human decision making redundant rather we see it as yet another tool to help humans make better decisions and probably make them faster incorporate more information but there's gonna be still a human in the loop a very good example for one of these systems is actually a project that we recently concluded um with one of our industry Partners in the Fashion retail industry I don't want to go into too much detail here but they were basically facing the problem that they wanted to enable something like one to two hour delivery from their store Network in major cities around the world in our case we've mostly focused on New York and that's a perfect example for something that many retailers go through right now kind of this transition from being a pure brick and mortar retailer to becoming an omnichannel retailer so actively opening the online Channel as a sales and distribution Channel and trying to basically integrate these two channels both the physical and kind of the digital channel and as we are talking now no longer about delivery speeds like two or three days but we're actually talking about on-demand delivery so in this case we're looking at a customer placing an order now and US guaranteeing that that order would get delivered within the next two hours so here we're obviously talking as I said before over highly fragmented distribution system where inventory can no longer live like 54 50 60 miles outside of the city but inventory has to live very close to where the consumer is and the question is obviously how do you design these networks how do you decide in this particular example which stores within the city of New York to activate for this service which capacities in the back room of these stores to actually Reserve not for the brick and mortar like the working customers but for the online customers of this company which inventories to potentially share between those two channels and how to dynamically adjust these allocations over time as for instance there's a a change in Market Behavior there is a holiday coming up and therefore demand patterns changing dramatically and in a situation like this traditional last month distribution Network design problems and methods typically fail because they make very strong assumptions about the problem being linear for instance they're not being any non-linear effects which is true if you are in the good old centralized distribution world but yeah it's no longer true if you're trying to optimize over highly fragmented Network because then at every single store you're facing highly non-linear effects for instance related to inventory but also related to capacity planning think of a store that gets too many orders that basically pile up in a queue so you're basically having a queuing problem and modeling these linearly is is not that easy anymore so um what we did here is we kind of For the First Time combined our traditional tool set from the or world with a new tool set from the machine learning world so we basically said okay there's no way we can model this problem to the level of accuracy that we would need by just relying on uh or model so on an optimization model so we need to make very radical simplifications to to make this work on the or side so we built a highly simplified optimization model that would solve the model that would solve the problem very roughly so that would give us a solution to the problem but a solution that would probably not perform very well because we had to make very strong assumptions that would not hold true in reality for instance we would have to just assume that linear that inventory requirements and capacity requirements at the stores would behave linearly in the amount of demand faced by these stores which is not true in reality so we use the simplified or model to generate Solutions and then we basically plug that into a very detailed simulator so we built a very defined grain simulation model that would actually simulate the city of New York in this case for the duration of an entire day we would show how orders come in dynamically how they would get allocated to the stores how they would get fulfilled so picked packed and shipped in these stores how The Courier Vehicles would move through the city serve the customer and throughout this entire simulation we obviously generated a lot of data we've generated a lot of data about how we think this type of solution would perform in the real world and that data then also reveals where our optimization model was off so we for instance could see that in a certain store our optimization model just allocated too few people for the picking and packing of online orders so we established too little capacity so and that's where machine learning came into play we basically built a relatively basic machine learning model that would identify these bottlenecks in the simulation data and would basically identify for instance which stores would need to get a higher allocation of capacity or higher allocation of inventory of a certain product and then based on these insights the machine learning model would suggest new parameters for our optimization model which we could then run again to get a new solution which would hopefully already perform a little bit better in real world than the previous one and so this kind of iterates it's an iterative process optimization doing its thing stimulating and then learning from the simulation so basically combining all three methods into one big model that then eventually helped us inform this company as to um yeah how they should design their last mile distribution networks for this particular highly challenging service um in New York City so this is a very good example of a relatively basic use of machine learning in our space the more kind of I would say Cutting Edge approach to machine learning or AI in our industry is actually concerning routing and this is where it's getting a little bit more interesting for us because what I showed you before the network design thing that is kind of machine learning alongside optimization and that has been done by now several times it's well established in the literature it's still a great method but I wouldn't call this the The Cutting Edge anymore The Cutting Edge really is trying to get rid of optimization or our methods all together in designing last night Logistics systems and honestly when we talk about designing last month systems we're very much talking about solving combinatorial optimization problems and the perfect example for a combinatorial optimization problem is the routing problem the vehicle routing problem there's Decades of research on how to optimally solve vehicle routing problems and honestly state-of-the-art methods from the or our world are pretty good so it's really challenging to basically set out and say okay we want to solve the vehicle routing problem but without optimization we want to do it with a machine learning based approach and we want to basically achieve competitive results from from that method um and this is ongoing research so this is very much work in progress right now but one of my students currently working on this in depth and we kind of realize well in a way if you think about what what it really means to navigate on a map so to find a good route that connects a bunch of customers it very much has a lot of similarities with learning how to speak basically learning kind of the inherent grammar of what a good route or bad route sequence looks like is kind of similar to learning the inherent grammar of how you basically find the right order to work to form a meaningful sentence that's on a 30 000 input level why we were we started looking into Transformer models in particular and other large language other methods that are mostly known to people from the large language model world um to use them for routing and um so the idea here was well some of the features that give these methods kind of an advantage when it comes to natural language processing might also enable them to actually solve combinatorial optimization problems such as routing problems quite well one is kind of the concept of having self-attention as I said I don't wanna end in this case I also probably can't go into a lot of technical detail but um self-attention is a mechanism that actually helps us consider the relationships between stops so for instance how far apart are two different stops or what's the traffic looking like between these two stops um positional encoding basically uh making sense of kind of the the sequence in which stocks are visited and in like basically incorporating that order information of the stops so basically we move from stop a to stop B to stop C incorporating that into the solution process and then having a layered architecture that is able to capture more complex relationships between stops as we move from layer to layer a lot of these properties that are very helpful in language processing are also quite helpful in combinatory optimization and as I said this is work in progress we're very very early stage in this but we're seeing some promising results here and um hope that we can actually kind of present some more details on this probably sometime later this year either the usual conferences or through an academic publication um the second thing that I wanted to talk about briefly in this talk is potential of drones and other autonomous technology investment Logistics and similar to AI we again don't think that there's really that one vehicle technology out there that has the potential to fully replace or substitute existing Solutions but a lot of our research actually suggests that it's more like a complementary element to Modern last modernistic systems so let's look into the example of drones in a little bit more detail and if you think about let's say drone delivery at first everybody is excited about it because it's just cool to have drawn slang through the air autonomously Landing in front of your house dropping off your Amazon packages and kind of conceptually that's a neat idea but when you kind of think a little bit harder about it at some point reality sets in and you realize well a flying drones is really hard from a technological point of view but apart from the technology side of things the even harder thing to do is to fly them legally and to fly them in a way that you're actually cost competitive so um we've been working quite a few years now on this topic with folks like ups and others and the bottom line of our research is the regulatory hurdles that are associated with doing commercial drone delivery to the end consumer at scale or just extremely high right now in most markets both in North America and Europe there isn't even a clear regulatory framework for this kind of service yet so a lot of this is still kind of in the process of being written and that obviously adds a lot of uncertainty as to when and if we will ever even be able to see this happen at scale um but apart from all the regulatory concerns that are out there the even bigger question to us is this really a desirable service from both the social and an environmental point of view and probably in both cases the answer is no one because if you're thinking about a highly fragmented consumer base and most of e-commerce volumes go to Consumers not to businesses so we're talking about individual package delivery to pretty much any household let's say in the US doing this at this scale would mean we have hundreds of thousands of drones flying around at any point in time which is hard to control but also not necessarily a nice thing to think about if you want to maintain the livability of cities it's not really desirable to have this guy buzzing of drones apart from all the safety related issues that might create but also from an environmental point of view moving things up in the air keeping them there moving them to your customers through the air is the most energy intensive way that you could possibly deliver a package so unless you are in full control of where the energy for that actually comes from and at least today we aren't and this is also not an environmentally friendly solution even though it might look like one but probably for most of the companies that think about this most striking argument why this is not a solution at scale is actually unit economics and again this is kind of on a very high level memorized from one of our research projects we looked at well what would be the unit economics of the most basic drone delivery model out there where you would basically just take a personal Center and equip it with an arbitrary number of drones and fly Parcels to customers from the parcel center and we we simulated various different uh scenarios here we looked into different possible realizations of the regulatory framework so right now for instance we're in the bottom right of this chart so we are today and we are in a very severe regulatory environment and so we we relax the regulatory constraints on the problem and we looked into the future to The Examiner we could and made some kind of projections about well how is technology going to evolve so for instance how is autonomy going to evolve how are component costs going to evolve in the next 5 10 15 years and for all of these scenarios we run very detailed drone routing models with a very realistic energy consumption module inside of them for instance to really understand the detailed cost of what it would take to deliver a standard e-commerce package to your door and what you see here is basically three things first of all we can basically already exclude the left side of this chart because we will never be living in a world where there is no regulatory constraints or just minimal regulatory constraints and I say never quite confidently here because this serves just too dangerous to not regulate it because if this thing falls out of the sky it hurts someone really badly so we will always be in that moderate to severe Spectrum and then if you look at the time Horizon here within the next five to ten years we will always end up in the death zone we will always end up in a Zone where unit economics are Way Beyond what they are with traditional ground-based delivery models so it really only gets economically interesting in the time already Rising between 10 and 15 years and that's already a kind of timeline where there's also so much uncertainty around how different component costs and the like might evolve that these predictions are taken are to be taken with a very big grain of salt so bottom line is at least in the foreseeable future we just don't see this happen at scale what our research suggests could happen at scale though is again models where we don't see the drones as kind of a pure play solution that goes from the origin of the shipment all the way to the final recipient so to the consumer in our case but where the drones play more of a complementary role to existing delivery models and one way this could look like is for instance enabling companies like UPS FedEx you name it to make the most of the assets that they currently have because these companies have buried millions of dollars into physical infrastructure for the type of delivery model that they're running today which is basically a hub and Spock system and in that Hub and Spock system everything runs kind of on a periodic schedule so if you if you want to know what that means is if you are for instance with UPS you will see that every morning I don't know between six and eight am at any given parcel center throughout the country you will see a huge line of UPS trucks lined up and Parcels are being sorted and loaded into these trucks they leave by I don't know 8 39 a.m in the morning and once they're gone they're gone and they're not going to come back to that parcel center anytime before the late afternoon what that means is that system is currently not equipped to support things like stop same day delivery so if there is for whatever reason a significant volume of parcels coming in let's say only at 11AM or at 1 pm and these puzzles would need to be delivered on the same day with this current infrastructure and system these companies won't be able to do it the only way to do it would be to basically put them on dedicated vehicles and send these dedicated vehicles in as kind of rushed courier delivery which is possible but a not very efficient not very environmentally desirable and also extremely costly so here the question is could drones actually help us leverage these existing systems and even though they are somewhat static um having spoke architectures can reuse drones to create more flexibility so a way to think about this is for instance using drones as a resupply vehicle to ground Vehicles they're already doing their tour so here you see an illustration of that where a ground vehicle leaves the Depot in the morning is doing its thing delivering to customers and then there are certain dedicated locations in the field in the field and these are denoted as trans shipment points here that the vehicle so the groundbreak can visit and meet with a drone that would resupply it with additional Parcels so Parcels that came in later throughout the day dynamically for other customer groups for other service speeds and it would pick them up from the Drone and then continue its route with basically an extended effective capacity because suddenly it had got an intra route resupply of additional volumes and that is a model that is a serving a real need it is basically providing more flexibility to the established distribution infrastructure and the established distribution model of pretty much all current parcel delivery services and it is making the use of drone technology more realistic also from a regulatory point of view because suddenly you're not just a serving kind of any random consumer that could be anywhere in the field but you're serving clearly defined routes from a Depot to a clearly defined set of Transportation points so you're basically you're flying the same Mission over and over again which is not easy but to get approval for but much easier than getting approval for individual deliveries to individual consumers um I'm sorry my cold is kicking in but um another way to think about drones as a potential interesting technology in this field is um looking at larger vertical takeoff and Landing Vehicles let's call it cargo drones for now they don't just carry one two or maybe a handful of parcels but actually a large number of parcels so think of the ideal case where you could basically put the entire load of a FedEx truck into the cargo compartment of one of these drones and that drone would actually fly from a parcel center somewhere outside of a city into the busiest areas of the city and meet the ground delivery vehicle there why is that interesting well if you look at the cities like Manhattan for instance you will see that many parcel delivery services right now serve up to spend more than 30 percent of their productive time every day just in traffic getting into and out of their service area so for instance a car that needs to go from the person Center outside of Manhattan into the busiest part of Manhattan can spend up to 30 percent of its route time just getting from the depot to the service center it's obviously lost time lost productivity and that's where drone technology could again be much more useful because it would address an immediate pain point it could overcome this highly congested inbound and outbound leg and allow the actually productive units so the delivery vehicle on the ground to start being productive not just at 10 30 am by the time the ground people would have finally reached its surface area but at 8 30 am so um allowing for better customer better delivery reliability better delivery speeds and honestly also just lower cost because of there being more time available for that after to be productive so this is kind of our high level view on these two things once on the technology side problems and other autonomous technology could be could be a way to design the future lost my Logistics systems could be the steroids that these systems need to basically keep up with what consumers need but not as a pure play solution not as a single technological solution that does it all but rather as something that complements existing systems and same thing AI has a lot of potential on the algorithmic side Ai and machine learning might be what we really need in order to even be able to design the and also manage future last analogistic systems but not as a standalone solution but always tying in the human and that's kind of the last couple of slides that I wanted to talk to in this uh talk about in this session um because my other hand at MIT is I'm leading the the cave lab the computation analytics visualization education lab where we work with people in a very visual way we basically connect all these fancy machine learning and optimization models that we built with human decision makers from our industry Partners through visualization and one of the main reasons we do this is because we feel like the more powerful the algorithms get that we built the more powerful the models get that we've built and the more data hungry these models and algorithms become the more there's a disconnect between what these models are actually able to do and what humans are able to appreciate there's a disconnect between the potential of these models and methods and the human ability to make use of that potential and we're trying to bridge that Gap through basically providing a better interface between the two making it more natural for humans to interact with Advanced Analytical tools and the reason why we do this is because Human Experience is important Human Experience in the supply chain Logistics domain cannot just be replaced cannot just be encoded in a model but there will always be a certain element of context awareness and experience by people who have been in this industry for decades that would be valuable to keep as part of the decision making process so basically to keep the human in the loop of the algorithm and that's where in this case visual interfaces are very helpful but um which also basically gives rise to my general call to the logistics industry but also to the research community that when we think about the future of Last Mile and research in this space we shouldn't just focus on one type of method we shouldn't just focus on one type of Technology but the real deal is the connection between all of this so for instance on the methodological side yes we can go on for another few decades just thinking about or methods and making them faster and better and I don't know what we can also jump full into the machine learning deep learning AI world and say okay that's where most of the really interesting research work is happening right now so let's focus on this but by doing either of these two things we would actually lose track of solving the real problem and solving the problem that real human and real companies out there actually have so we think research is very much focused on hybrid methods how can we combine The Best of Both Worlds the our world and the Deep learning world um similarly how can we keep the Union in the loop how can we design algorithms that are very powerful but don't just fully replace human decision making but actually bring in that Human Experience give the human kind of the ability to give feedback to an algorithm give feedback to a model evaluate model Solutions naturally very quickly and basically be part of this decision finding process and here let's say for instance the cave lab is already doing quite a lot of work on the UI uxi so mostly finding visual cues that help a human interact with an or model and we think in the future we want to put a little bit more effort on actually the the language side of things so that becomes mostly important when we want the human to also be able to more naturally interact with artificial intelligence or machine learning models um visualization doesn't really help you with that you really need to be able to have like a language connection between the human and the model think of chat GPT that's the perfect example for for such a language-based connection and so there's a lot of really interesting methodological work that could go into designing the future lost my Logistics systems and um yeah that's why personally I am very excited about the next couple of years of research that we're going to do at the lab and that MIT as a whole and we are obviously hoping that this also finds enough attention in industry and with our academic community to basically form the right projects form the right teams and and get this research off the ground and with that I think I'm done and I probably went a little bit over time but this was like my general view on where last man Logistics might go in the future thank you thank you so much Matthias that was that was a great talk I I want to start by just uh double clicking on your final Trend which was sustainability it's not just lip service anymore so interesting I was I was watching a talk from an executive at Amazon and she said look in the United States transportation is about one-third of greenhouse gas emissions and Logistics Goods Mobility is about one-third of that and last mile is about one third of that so by you know the relation what they're doing with their project with rivien to Electrify the delivery Fleet for Amazon you know will have a real impact now that's that's on the supply side side what I want to ask you about is the mobility demand side so the question is really about you know this instant gratification Trend that we've seen um and do you see that we could potentially price that you know so we could implement the the retailers have you seen movement among the retailers to effectively price in well look if you want to order your toothpaste at eight o'clock in the morning and then you remember oh wait I need mouthwash at noon and then just before bedtime you remember oh and by the way I needed a toothbrush and put in these individual orders which of course you know further put further Demand on the system is there a way to price that number one and do you think that consumers are willing to pay for that number two um to answer number one yes there are ways to price that um the question though is whether anyone wants to price it because I mean even if the consumers would be willing to pay for it and buy now I would be skeptical about that because let's say we're all used to getting it for free or more or less for free by now and I don't think that it's going to be super popular to suddenly start charging for it again but even if people were willing to pay for pay for it let's assume that for a moment I don't think there's going to be a lot of companies who would want to charge for it for the very simple reason that right now this ability to provide that level of service that level of flexibility that level of speed is kind of the key differentiating factor between some of the big players like most notably Amazon and the long tail of the rest well the it's a it's an entry determinant mechanism mechanism basically because if you do not have the volume and the scale of someone like Amazon or maybe Walmart or I don't know you can't do this cost effectively because this is pure scale game if like for instance 2019 I think it was nightly announced okay we're going to stop selling directly on Amazon we're going to do we're going to double down on our own online Channel and yeah from a branding point of view makes total sense but from a logistics point of view that's a nightmare because even the huge Brands like Nike will not have the volume and that means will not have the density of customers that would allow them to keep up with these types of service levels in a way that's even remotely cost effectively so um long answer to your question but I think yes it should be relatively easy to price this but there's not a lot of willingness to prices neither from the consumer side nor from the retailers yeah yeah well hopefully a starting point on that is just raising awareness among consumers about the the consequences of fractions I forgot to mention that so we did do well not my team but one of our other teams at TTL did it w three years ago with a retailer in Mexico where they actually experimented in their life system so on their online store um whether people will be willing to go for slower delivery service if they not had to pay for it uh if so they didn't charge for the faster service but they basically said okay if you go for the lower service that's going to save 100 trees and so they made it very tangible but they didn't monetize it and that was actually interesting to see that was quite an effective lever to make people aware of the environmental impact and reconsider well apart from my own financial situation or the economic impact that this might have what am I doing to the environment by this and do I really need this service so that is I think the more effective way to think about it I have one more quick question before turning it over to move on and it was related to your point about the proliferation of micro fulfillment centers and cities and how do we position Last Mile facilities storage within within cities I my question is you know if you had you know 10 City Mayors in the room asking you for advice or perhaps a real estate developers what what would you tell them what would you tell them about the fabric of the city more from an Urban Design and planning perspective to look for going forward into the future yeah I mean I'm not an urban planner or a real estate government so they might have different answers to this but if I was talking to a city I would basically ask them to think about the current land use within that City and think about the spaces that they think are currently generating the least amount of Revenue per square foot basically so what is the basically least value generating used right now of space in your city and that's literally for any space not commercial space like any base speed office based apartments parking lots you name it because these spaces regardless of their nature will probably be the ones that are most likely going to be replaced by Logistics activities in the future because these microfulfillment centers I mean in an ideal world you would obviously want to build a tiny little highly automated Warehouse from scratch but in most densely populated cities that's not a reality you do not find an empty lot in the middle of Manhattan that is just waiting to become the warehouse that's not happening so it's a lot about repurposing existing space repurposing parking garages for instance but also quite frankly repurposing space that might have previously been a retail store or that might have previously been an office now that everyone's working from home or in the most unfortunate event might have previously been an apartment so if cities want to understand the impact of these microphone centers or like hyper local fulfillment in general they should think about which spaces are the most likely to be replaced and do I want that from a political point of view from a societal point of view so if it's for instance the apartments that might be a problem if it's parking lots that might actually be desirable and so um I think from a policy point of view that's the most interesting angle to this super we I'm going to turn it over to move on we've got a ton of questions on sustainability as expected so John touched on you know the carbon pricing and you mentioned you know companies are a bit hesitant even though you know consumers may want to pay for it uh so there's a question from Ryan Westrom so he says if companies are not willing to price in the costs of inefficient ordering or delivery is there a needed role for policy regulation because how else do we get sustainability you know if everybody's just ordering ad hocally you know companies are subsidizing this and you know just wanting more consumption so you know how do we make this more sustainable you know that it I know it's a it's a loaded question but yeah I mean it's more of a personal opinion but from all I've seen in my work in this field so far regulation usually doesn't fix the problem in many cases cases regulation just makes it worse and so I think the question is rather how can we accept the fact that humans are imperfect and that humans do order the toilet paper five minutes after they placed their other Amazon orders because they just realized they're running out of toilet paper so this is just human behavior that's not going to go away whether you penalize it or not um so the question rather is how can we make the logistics systems that have to deal with this more flexible and obviously a lot of that has to come from the companies who operate these systems so that's why to stick with this example folks like Amazon and others invest a lot of people a lot of money into coming up with better algorithms to actually predict this type of behavior so that even though you as a individual consumer might have a completely erratic ordering pattern they can somehow anticipate that and pre-position inventory more effectively such that the environmental impact of that might be limited so for instance such that even though you place five different orders throughout the day you still only get the one delivery that's what the companies need to do on the algorithmic side quite frankly but a city could for instance think about well how can we provide the right infrastructure or more flexible and more sustainable delivery options so a big question Topic in this space for instance unattended delivery because one of the main reasons for highly fragmented deliveries right now is that most of the delivery still go to the individual home to the individual address and that comes with its own challenges so can we create spaces within a city where we can do safe but unattended deliveries safe and unintended kind of pickups to basically decouple the availability of the consumer from the actual delivery process because that will then enable the companies again to more easily consolidate loads consolidate shipments consolidate routes because you're relaxing some of the constraints on their routing problem basically more potentially in enabling rather than restricted so I get the bind that Consolidated deliveries can you know help it make more sustainable so let me pose the question of food deliveries you know there's this bad image of you know a whole car moving around with a small package of food to be delivered to somebody in the city so uh this is the question from Michael Leon so you know where is the consolidation happening there you know it's spending so much time in urban traffic you know they are double parking all over you know Copley Square in Boston and uh so and and this is this is an increasing Trend where it's it's getting worse with you know as time progresses so and you said you worked on food delivery as well so where does the question of Consolidated delivery come in you know where in in the case of poodle yeah and we worked on food delivery in an Indian context which is from a traffic point of view even crazier than anything that we know in the US for instance yes more and so much more congested um and here I mean again I think it's very hard to change consumer Behavior I don't think regulator even if they wanted to could keep people from ordering stuff on I don't know ubereats or daughters or I don't know how they are called so um you're not going to change the demand side that easily unless you impose very heavy economic disincentives and that's hard to do at least in the market like the US probably even impossible so the question really is what what needs to happen to keep up with this demand but require fewer resources so again a lot of it goes and that's what we worked on with the Indian food delivery Network for instance how can we make these deliveries more efficient how can we build better algorithms so that it's not just this one order going around on a single vehicle but maybe four or five so how can we connect multiple orders multiple customers and in this case also multiple restaurants on a somewhat Consolidated route obviously the potential for improvement there is limited because of the heavy time constraints on this but it's an important lever because right now we see too many cases where it's still just the one order per vehicle and you can definitely do better than that the other side again is about um infrastructure so for instance I think Boston is a great example for this um Boston is investing relatively heavily in safe bike infrastructure biking infrastructure for instance because for food delivery networks in particular if they can take it off a motorized vehicle so if it's not a car or a van or whatever they will because bikes will always be more efficient in doing this than cars the reason why it's sometimes not possible is there's not enough safe infrastructure for people to do this on bikes or it's just not a very bikeable City in general so um again that's where probably a city could say more okay how do we accommodate this trend by creating sufficient and safe infrastructure to to make this less disruptive and also polluting um while we will not change consumer Behavior yeah Matthias Matthias when came back the head of the MIT Mega City Logistics lab and the MIT cave lab uh thank you so much for your time today Matthias we're going to share the chat you know Robin Chase for example had some very interesting comments uh I always find the chat file very interesting to read but thank you so much for your for your contribution today thank you sure

2023-05-05 01:10

Other news