Amazon Robotics’ Dylan Randle met with students from the Break Through Tech AI program to discuss the practical applications of machine learning and robotics at Amazon.

Liz McShane
August 11, 2022

After a studious few weeks of lectures and practical labs on machine learning, participants of the
Break Through Tech AI program interacted with professionals in a series of industry-led lectures. Organized by the MIT Stephen A. Schwarzman College of Computing, this year’s program is hosting 33 undergraduate students from colleges and universities in the Boston area.

On July 12th, students gathered in an MIT classroom for a lecture by guest speaker Dylan Randle, a data scientist from Amazon Robotics. At Amazon, Randle analyzes ways to make the order fulfillment process in their warehouses more efficient—which can range from ensuring a fleet of package-carrying robots can move around without bumping into each other, to making sure that robotic arms handle packages with care.

Randle shared details about the challenging tasks his team is grappling with and students punctuated the lecture with incisive questions. Data scientists at Amazon deal with a complex decision-making process: packages need to be inventoried, stored, selected for picking and packaging, moved and shipped. Randle’s presentation exposed students to real-world industry challenges that need to be solved.

There are three major steps to order fulfillment at Amazon. When buyers place an order, their item is directed from one of the massive warehouse centers to what Amazon calls the Middle Mile, which includes a transportation network and a center that sorts packages. From there, the items reach a delivery station, where the packages are transported to their next destination. Amazon Robotics’ goal is to make this process more autonomous and efficient at every step. Randle explained how computer scientists and engineering teams apply state of the art machine learning at various steps of the order fulfillment process to solve the many logistical challenges.

The first step takes place in enormous warehouses filled with robotic pods that are stacked with the items Amazon offers. 

“How are items organized at the fulfillment centers?” a student wondered. “Is there a way to speed up the process?”

“We use machine learning a lot!” Randle replied, “Amazon is investing in new algorithms to sort packages and organize goods, but this is an on-going challenge. Applied machine learning algorithms that could reduce the timing of order distribution from hours to minutes are being tested every day as we search for scalable solutions.”

A student who was virtually participating asked “What task would a software developer be responsible for at Amazon? Would it be maintaining code?” 

“Software engineering is one of the most versatile jobs,” Randle noted, “Coding skills could translate to a wide array of job opportunities at Amazon.” Software engineers are responsible for building Amazon’s systems, including maintaining software code and creating new softwares. They might also build new tools or even robots. 

Opportunities to ask questions like this helps students understand the real-world scope of jobs in the fields of data science, machine learning and artificial intelligence. The Break Through Tech AI program at the college prepares women and students from underrepresented identities with a competitive portfolio and technical skills that ready them for industry jobs after graduation. The program is a DEI initiative with a greater mission: creating a more equitable job market. Guest speakers like Dylan Randle share professional experiences of what their daily jobs entail and highlight the importance of diversity, equity and inclusion in the tech ecosystem.

Randle emphasized the practical nature of the problems that need to be solved at Amazon. Robotic manipulation is one field that is “near and dear” to Randle, so he elaborated by talking about a new robot being used in the Middle Mile to speed up package sortation. He described it as “a giant arm with suction cups” that are used to pick up and move packages.  

This robotic arm faces many physical challenges: It has to sort, lift, seize, and move packages of different shapes, weights and materials. Unveiling the practical issues that industries are dealing with on a day-to-day basis is a key part of the Break Through Tech AI training. 

“Does this robotic arm have the capability to detect that it has picked up the wrong object?” asked another student.

The robots currently do not, and Randle described a “worst case scenario” in which a robotic arm could create a “traffic jam” by picking up a package and accidentally dropping it in an aisle. “This does happen and it is a key situation we have to deal with in real time,” Randle said. 

In this scenario there is a waterfall effect, as the the robo-arm is unable to move onto the next package, the item that was dropped blocks the smaller, transportation robots that need to drive through the aisles, and there could be shipment delays if the package doesn’t get picked up. “We have to develop new solutions to minimize these ‘traffic jams’,” said Randle.

Amazon Robotics is using simulations to help solve real-world case scenarios like this. “Simulations created in virtual reality and graphics software are key to improving robustness by training the robot with many more possible or problematic scenarios, including events that can be very rare but still plausible in a warehouse,” said Randle. 

At the forefront of this simulation software is a program developed at MIT called Drake. It is an open-source toolbox created by the Robot Locomotion Group at the MIT Computer Science and Artificial Intelligence Lab that analyzes the dynamics of robot movement and builds control systems for them.  

“Drake models forces and deformation in grasping objects,” Randle said, “Simulations are getting so accurate that you can generate millions of simulated data points to augment the dataset an ML model will need to learn well.” Randle is excited about the possibilities simulations provide. “This is a very promising approach that helps to prevent mistakes, as it trains the ‘brain’ of the robot to be ready with cases in the real world that it may encounter only one time,” he said.

Robots are currently controlled by a centralized path planning program that determines where they can go in a warehouse. Introducing new robots to the order fulfillment process adds flexibility, but also complicates path planning. Randle notes that with a fleet of robots all moving at the same time along the many aisles, there are “more possible paths than atoms in the universe.”

“The biggest issue with these robot drivers is that we end up getting congestion because drivers can’t go through each other, or jump over each other.” This requires teams of software engineers to plan optimal, collision-free paths, where robots are kept in motion, without colliding or having to stop. “It is a very important job for us,” Randle adds, “The software engineer team has to program a fleet of robots to navigate safely and make decisions on the fly.” Like bees, each mobile robot has a task to perform, but they are also working in concert. 

“With a centralized path planner, you have guarantees about exactly what will happen and how the system will behave, but it also becomes a single bottleneck for everything. If you had a system where each robot made its own decentralized decision, it may be more efficient but you would not have guarantees about what was going to happen when the entire fleet is moving. Though decentralizing the system would mean you wouldn’t have to have servers computing many paths, it’s not as predictable.”

While Amazon’s pathfinder system is currently fully centralized, Randle sees the future moving towards more independent robots that have decision-making abilities for autonomous drives.

At the end of his presentation, Randle talked with the students about internships and job opportunities at Amazon, inviting them to further connect with Amazon Robotics. Industry guests extending a hand like this hits home the mission of the Break Through Tech AI program. The students were able to learn about practical issues in automation and take their education beyond the classroom by being exposed to real-world applications of machine learning.