From Autonomous Agents to Travel Agents: Applying Lessons from Robotics to Travel Planning

I studied robotics during both my undergrad and masters degrees at MIT. Now that I’m out of school, people who knew me in my graduate program often ask what I’m doing for work, and people at work often ask what I researched at school. Generally, people are surprised to learn I’m now on a team at Mobi that focuses mostly on travel and logistics software. “How did you decide to make that pivot?” they ask, and I usually respond that actually—it wasn’t a pivot at all! In fact, all of Mobi’s planning algorithm teams (planner teams) employ quite a few robotics PhDs. The tools we learned in our academic programs that help us schedule and coordinate robots turn out to be the same types of tools we use at Mobi to schedule and coordinate humans.

Some members of Mobi's team with robotics backgrounds:

Mobi’s planner teams use classical AI techniques to generate schedules for the transportation and hospitality industries. Usually, when I summarize my work this way, I put stress on the “classical AI” part to explain that, as opposed to using a different AI approach like a large language model (LLM), we aren’t doing routing and scheduling with an AI-chatbot or throwing a bunch of machine learning (ML) techniques at a highly constrained problem to see what sticks… but in this technical piece I want to focus on the second half of this statement—what does it even mean “to generate schedules for the transportation and hospitality industries”?

Individual travelers often want to make the most of their week-long getaway to a foreign country. The upfront cost of travel and the preciousness of vacation time make even those who are strongly-opposed to scheduling their lives consider an itinerary–whether it’s scheduling down to the minute to catch a ferry or deciding whether you’re dedicating a specific day to the beach. Planning for your vacation gives something for people to look forward to and assures them that they are using their time effectively. You may have made one of these schedules yourself, gone on a scheduled/planned tour, or worked with a travel concierge to make one (or maybe your partner or family member has lovingly made one for you before!). Even if you haven’t, you can probably imagine what one would look like: activities you want to experience, food you want to eat, logistics, and when you want to sleep are shuffled around until they form a schedule of locations and movements, complete with buffer time and perhaps granular to the day, hour, 15 minute, or one minute interval depending on the required precision. Developing an itinerary becomes even more essential when you’re coordinating multiple people for a vacation. What time and for how many people and where do we make a dinner reservation for Saturday night? Can we all make it back to our different hotels to change before going dancing? How long does it take to get between the hotels anyways? We can also think about schedules coordinating even greater numbers of people: with knowledge of when and how tourists want to arrive, we can generate better bus, shuttle, train, and plane schedules to maximize efficiency and comfort. These schedules look like travel times + routes, combined with meet up location + times.

Everything I just described to you may seem incredibly commonplace and intuitive–I’m just explaining how one would pre-plan a trip or load balance transportation–but now let me explain how this is actually quite similar to my previous research in robotics. When pre-planning for a series of tasks–a “mission” let’s say–undertaken by a robot, there are many ways to write down “the plan” you’re going to feed to the robot. One of the common approaches is to generate a location-aware chronological schedule. For example, if I am generating a “plan” for a Planetary Rover, where the rover, the mars rocks it needs to collect, the astronauts, and other robots all need to be at specific times (down to say, an hour, 15 minute, or 1 minute intervals) is a location-aware schedule. The rover’s computer can easily “execute” this plan as long as its “fine-motor skills” are executed by other software. You could even export it to Google Calendar! Every “agent” on the team (the astronauts, the ground-support team, the other robots) can use the same planning engine to receive their own, coordinated, location-aware schedules.

Is it a coincidence that one of the best ways to schedule and coordinate robots is so similar to scheduling and coordinating humans? Not at all; location-awareness and the timescales on which the robot-schedules operate are driven by the robots being embodied. Being embodied here means robots–like humans–must be aware of their location/position, are similarly constrained by the laws of physics in their movement, and are often executing tasks at speeds humans can interact with–or at the very least can understand on human scales. By contrast, scheduling for “unembodied” computers (scheduling for cloud computing or operating systems, for example) aren’t normally location-aware because the “location” of a task doesn’t matter, a task can be done anywhere, or sometimes a task is done in multiple places at the same time. “Unembodied” computers also operate on much faster and less understandable time scales. For example, schedules for operating systems might switch between tasks so fast humans don’t notice, or can start and abandon a task multiple times because restarting is faster than finding where the old task left off. If I had been working on schedule-generation algorithms for operating systems instead of robotics, it would have had much less overlap with travel planning!

But generating a “location-aware itinerary” for travel using the same algorithms that are used to generate plans for robots isn’t the end of the story. These similarities are why, when Mobi looks to the future of transportation and travel planning, we see so much potential for growth. The field of robotics has so many other planning algorithms with untapped potential. Mobi is bringing these algorithms over to other contexts. This is also why Mobi thinks planning your vacation could be so much more than the travel-industry standard “print-out of your pdf itinerary before you leave”; because we’ve seen the technology that can make dynamic, coordinated, location-informed itineraries in other contexts.

A few examples of applying robotics lessons to logistics and travel problems:

When plans inevitably go wrong, the field of robotics has “live” algorithms that make adjustments to your plans as you travel. Imagine missing your train from Milan to Venice and your travel app automatically regenerates a new itinerary: adjusting your transportation schedule, your dinner reservation time, and recommending a new activity.

The field of robotics also has systems for coordinating “multiple agents”, which can be applied to coordinating the plans of multiple people. Maybe, going back to the last example, you were also meeting a friend in Venice when you missed your train. Using team-coordination robotics algorithms, we can also change your friend’s itinerary in real time when you miss your train so you rendezvous at the correct time.

Robotics scheduling has concepts of “risk tolerance” that can be used to build in smart travel buffers. Instead of adding a flat “15 minute buffer” every time you move from location A to location B, we can adjust the length of the buffer based on the reliability of your transportation, how strict the start time is of your next activity, and how much being late or early affects the rest of your itinerary.

The field of robotics is also held to high standards of “safety” and “explainability.” If a robot is helping an astronaut assemble an instrument, we need guarantees the robot won’t suddenly move sharply and injure the worker. If a rover decides to take a right turn when you were expecting it to take a left, it is expected that the rover support team can look at data and the algorithms to “explain” why the rover made that decision. We are also able to use these “safety” guarantees and “explainability” algorithms to make our decisions about transportation and travel planning more transparent and understandable.

Additionally, the algorithms we are talking about can both “make guarantees” and “optimize” separately or together. For example, sometimes we need to know if it is possible to physically visit every museum on a traveler’s “favorites” list during their visit, and other times we want to know the most efficient order for routing between those museums. Sometimes we want to know if a traveler would even have time to ride every amusement park ride, and other times we are looking for the most efficient schedule to avoid lines. This can be expanded to other objectives beyond routing and scheduling, like personal preferences or making a budget. We additionally have algorithms from robotics that allow us to achieve multiple objectives at the same time or relate the objectives to each other. So I might say: “find an itinerary that meets my budget and time constraints and then optimizes my route for as little travel time as possible.”

Coming to Mobi has allowed me to take robotics—a discipline I loved learning about and working in—and apply it in an impactful way. I’m excited to continue revisiting algorithms I used in robotics in different contexts. Not only because doing this brings a fresh set of tools to travel and logistics, but also because adjusting these algorithms to new problems makes me think in novel and creative ways. By pulling from unexpected places, Mobi has created an environment where interdisciplinary solutions thrive.