calit2

We Welcome Our Robot Tour Guides

Take a walk inside the entrance to UC San Diego's Atkinson Hall and you might be greeted by a robot made of two cardboard boxes stacked one upon the other, complete with a signature smiley face. BoxBot, as the bot is called, works with his counterpart TritonBot, a robot with a fully functional body, in the hopes of one day becoming a lobby receptionist for the building. But the 'bots are also being used for more sophisticated purposes: to collect information about how humans respond to robots. 

A student interacts with BoxBot
A student interacts with BoxBot in QI's Atkinson Hall. Photo by Alex Matthews

The robots are “chaperoned” by several graduate students, including Computer Science and Engineering Ph.D. student Shengye Wang of the UC San Diego Contextual Robotics Institute, which is based in Atkinson Hall. Wang, along with postdoctoral researcher Heerin Lee, is working with Professor Henrik Christensen to create robots that have the long-term memory autonomy (the ability to remember faces and names, for example) required for interaction with humans. The two robots have been greeting guests in the lobby of Atkinson Hall for about a month, and will be part of QI outreach efforts on Friday, March 23 when they will greet visitors at the institute’s free monthly tour, from noon to 1 p.m. [Note: The tour is currently sold out, but QI will host another free tour in April.]

About the Bots
TritonBot is a mobile, research-edition robot designed by Fetch Robotics and is fully functional, with an upper body, three-dimensional camera and a robotic arm capable of seven degrees of freedom. BoxBot, on the other hand, is a “freight” robot from Fetch Robotics that is topped with boxes for a “torso” of sorts. Wang says the boxes are a means of quickly prototyping a shape and design for the upper body based on the ease at which people interact with BoxBot. BoxBot's "head" (its top box) is also equipped with a mounted camera, microphone and speaker.

The robots are programmed to detect people's’ faces as they enter the lobby. If the robot recognizes a person’s face, it says, “Nice to see you again” and if not, it introduces itself and asks the user trivia questions (example: "Where does the president of the United States live?"). While the person answers, the robot converts captured facial images to a series of numbers called embeddings and stores these numbers with the name of the user. It later “remembers” the face so it can say hello to the person by name at the next interaction. It recognizes faces by calculating the embeddings and comparing them to those it previously seen, finding the best match with which to greet people.

Fetch robot
A manufacturer's photo of Fetch Robot

Throughout the month of February the robots collected 400 gigabits of sensor data. Some of the data is redundant, and some information is missing. Wang says it will take a lot of effort to look through this data and summarize a lesson they can learn from the month-long deployment of the robots. 

"For the software and hardware failures," says Wang, "we are working on formalizing the research questions and building fault-injection/simulation tools that allow us to replicate the failures in TritonBot and BoxBot. As for human-interaction, we found that people tend to respond before the robot finishes talking, and we change the robot to start listening slightly before when the robot ends talking. Also, the robot failed to respond to human response sometimes, and we are looking into the transcript of all the dialogues to understand people's expectations of the robots. We will use our lessons learned from the first-month deployment to improve TritonBot."

Wang theorizes that an interesting potential future study would be to analyze the distance visitors to the building will follow the robot related to the age of the visitor (children, for example, might follow at a closer distance). Existing algorithms would allow the researchers to estimate a person's age using facial recognition technology, and the robot’s laser scanner could estimate the location of people’s legs. From that data, they could calculate the distance. 

In the meantime, BoxBot and TritonBot will continue to periodically greet visitors ... and (as is so often difficult for humans to do) try to remember their names. 

Related Links

Contextual Robotics Institute

Media Contacts

Tiffany Fox
(858) 246-0353
tfox@ucsd.edu