MIT Finally Does Some Useful Research With Beer Delivering Robots

A multi-robot planning algorithm makes sure that you and your friends get the beer deliveries that you need

3 min read

Evan Ackerman is IEEE Spectrum’s robotics editor.

MIT Finally Does Some Useful Research With Beer Delivering Robots
Photo: MIT

I have this suspicion that if it weren’t for beer, robotics research would be years, perhaps decades, behind where it is now. This is because beer is the answer to the question that everyone always asks about researchers about their robots, which is: “That’s cool, but what does it do?” If you can somehow answer that question with, “It brings me beer,” people immediately understand the value and importance of your research (and by extension robotics in general), no matter what it actually is.

Having said that, getting robots to deliver beer does in fact involve a lot of complex issues, including navigation, perception, grasping, and human-robot interaction. And once you’ve solved all of those, you still have to get groups of robots working together if you’re trying to deliver beer to all of your friends at the same time, which you totally should be. At MIT, they’ve developed a new kind of multi-robot task planner that enables beer deliveries to consistently occur even under uncertainty, and that’s a thing that the world obviously, desperately needs. 

The problem the MIT researchers are trying to solve is how to get teams of robots working together intelligently in situations where one robot might not have a very good idea of what another robot is doing: in other words, situations where communications might be unreliable, which includes most situations outside of the nuclear-powered Wi-Fi that probably exists in your lab.

To help robots collaborate under uncertainty (whether it’s related to communications, sensors, or outcomes), MIT has developed a new type of planning algorithm with maybe the worst name ever: a MacDec-POMDP planner. To deconstruct that, it’s a Decentralized Macro actions planner for Partially Observed Markov Decision Processes. To deconstruct that even further, I’d recommend this excellent page on POMDPs for Dummies.

Anyway, this MacDec-POMDP planner enables individual robots to engage in high level reasoning about what’s going on and what they should be doing. In practical terms, this reasoning ability makes the robots much more capable of executing tasks like “get me a beer” if they’re not sure where other robots involved in the task are, or what they’re up to.

The key difference is the use of what is known as a finite state controller rather than a policy tree: while a policy tree can be thought of as a series of “if this thing happens, do this other thing” rules, a finite state controller is based on a high-level model of the problem to be solved, and it focuses on results, rather than details of execution. This makes it much easier for the robot to develop a plan based on different situations; in effect, the MacDec-POMDP planner can automatically generate controllers for the robots that maximize the utility and efficiency of a multi-robot team, even if each individual robot has a different amount of uncertainty that it’s dealing with. This is not an easy concept to really understand, but there’s lots more detail in the paper, linked below.

The video demonstrating the planner in action with a PR2 bartender and two TurtleBot waiters is a bit dry (so maybe find a robot to get you a beer to drink while you watch), but the underlying research is very cool:

This research does have other applications beyond beer delivery, as preposterous as that may sound. For example, in a warehouse or disaster environment, you probably have a bunch of robots all trying to coordinate to complete well defined tasks, probably without nearly as much information as they would like, and this kind of planner allows them to be consistent and efficient in their actions.

Just for fun, let’s take a look back at the original PR2 beer fetching video, which is now over five years old (!):

And where is the Beer Hackathon Team now?

  • Eitan Marder-Eppstein – Senior Software Engineer, Google
  • Radu Rusu – CEO, Fyusion
  • Melonee Wise – CEO, Fetch Robotics
  • Caroline Pantofaru – Senior Research Scientist, Google
  • Jonathan Bohren – Ph.D Candidate, Johns Hopkins University
  • Gil Jones – Software Engineer, Google
  • Curt Meyers – Engineering Guru, Suitable Technologies
  • Brian Gerkey – CEO, Open Source Robotics Foundation

That’s a kind of scarily impressive list, I must say, and a bunch of these people did other awesome things before ending up where they are. So let this be a lesson to you: teaching robots to deliver beer results in future greatness 100 percent of the time.

[ “Policy Search for Multi-Robot Coordination Under Uncertainty” (pdf) ] via [ MIT ]

The Conversation (0)