AMAZON ROBOTICS
Hardware Design Engineering Intern
Summer 2022
As part of the Sparrow team that was tasked with manipulating prepackaged items in crowded totes and isolating them for future sortation, I worked specifically on the end effector in order to reduce damages and develop a better testing method for future end effector comparisons. An overview of the Sparrow robot and its corresponding end effector Daisy can be seen below.
While the video highlights many of the cool features of Sparrow including multi-modal sensing and advanced perception I was specifically tasked with the hardware design of Daisy, the end effector. While super effective at picking some objects there were certain objects where damage was super common such as book covers, shoeboxes, mesh objects, shrink wrap, or things with pointy and unique shapes. However it was initially really hard to understand the pros and cons of each manipulation method because each test was run in a different environment (tote vs isolated vs crowded) using different test sets. It was through the first weeks of onboarding that I learned the importance of developing good tests and standardizing control variables during testing.
This inspired be to spend the first half of my internship going through the data on other end effectors in development to really understand exactly where Daisy was failing and if any other end effectors were able to solve that problem. As such a developed a new standardized test set and procedure, that was reusable and repeatable something that didn't exist before I started. There were two main tests - a Daisy current damage test set as well as an overall representation of the objects that a typical robot could expect to see (even though it varies greatly by time of year and location). This work then let into the design of a new hybrid end effector that build off Daisy's success, while not damaging even more of the objects.
​
One of the biggest challenges I encountered in this role was designing for scale. Having never worked on such a big project, I usually deprioritized cycle life and repeatability in my designs however when working on this team, I was not just thinking about parts per million but wanted to ensure that the end effector would be able to last multiple million picks per day for over 6 months. My new design combined a central toroidal gripper with 6 suction arranged similar to Daisy. This allowed the software to choose between manipulation methods depending on the circumstance and reduced damages by 60% according to my testing procedure. However the flexible membrane of the toroidal gripper was very susceptible to being punctured and stress over so many cycles which would need to be tested further before being deployed into fulfillment centers.
​
6 months after my internship Daisy was deployed over other end effectors in development. While a lot of the design is kept confidential and a lot of the internal testing is not available publicly, the video below, an artifact from my internship shows Daisy undergoing one of the tests. Also attached is the Amazon Manipulation Picking Dataset - using some of the data I collected and releasing it publicly to encourage research and development with these challenging manipulation challenges.