Elaine standing in the USC interaction lab, holding a small white robot.


My goal is to develop algorithms that enable robots to help people, especially people with disabilities, as they go about their daily lives.

Research Interests

Providing continuous physical and social support in human environments and social interactions requires new algorithmic approaches. My research enables service robots to make effective use of computation to address the most critical elements of interaction with humans, while being flexible enough to support the full richness of human behavior. This includes developing fast, data-efficient algorithms for group interaction in noisy real-world environments, algorithms for integrating task and social behavior over time, and understanding how algorithmic choices affect perceptions of robot agency

Google Scholar Page CV (pdf)

Efficient Interaction with Groups and Crowds

My work allows robots to influence, understand, and learn from people in groups, while making minimal assumptions about the specifics of their behavior. My goal is to enable robots to appropriately interact with and learn from people in public spaces, with learning algorithms that appropriately integrate information from diverse users and control algorithms that appropriately respond to human behavior in noisy environments.

The tall grey robot named Poli waves at a young woman in the atrium of an academic building.

Detecting Contingent Responses

This work developed a real-time algorithm for detecting whether the robot received a response to a greeting that matched offline performance (89.5% real-time vs 91% offline). Using a small training set (14 positive and 14 negative examples) this approach allowed the robot to choose fewer moments to ask someone to do a survey (13 instead of 38 in a naive approach) while more efficiently obtaining responses (30% response rate rather than 26.3%).

Published at HRI 2018


A group of people talk to each other and to the small fuzzy robot named Chili.

Developing Robot Moderators

This work developed one of the first algorithms for proactive robot moderation of collaborative group interactions and showed that a moderator that selected actions to support existing social dynamics could increase collaborative behavior by up to 20 episodes in a 6-minute interaction.

Published at Ro-Man 2014


A closeup of the short grey robot Moe using one arm to pick up a block from a table.

Learning from Crowds

This algorithm takes a single demonstration of a manipulation task and learns new approaches to the task by interacting with users in the wild. Using simulation, execution monitoring, and active learning, the robot autonomously performed 496 manipulation actions in a public space, collected 163 labels from users, and learned up to 49 new valid sequences to put the block into a bowl and 70 new sequences to pour it out

To appear in HRI 2019

Video PDF

Integrating Task and Social Behavior Over Time

Social interaction increases the importance of timing, since small changes can have social meaning (e.g., synchronization both results from and reinforces rapport). A core algorithmic challenge I address in my work is how to integrate socially appropriate behavior into temporal models for planning and control algorithms.

The small fuzzy robot Chili sits on a table among various food items, playing with a young girl of about five years of age.

Teaching Nutrition to Children

In one of the earliest long-term child-robot interaction studies, the robot used story-based social interaction in combination with a cueing algorithm to scaffold children's learning of nutrition information. In a 6-session study, we showed that this algorithm effectively combined social and task behavior: in the final week, children used more complex speech, responded more quickly to the robot's social questions, and took more time to respond to increasingly-difficult academic questions without losing performance.

Published at Ro-Man 2014


The tall grey robot Poli drops a cup while standing in the atrium of a campus building.

TPNs for Multimodal HRI

This compact architecture for orchestrating time-sensitive reactive social behavior with planned task behavior uses a variation on Time Petri Nets (TPNs). This representation is compact (with a graph up to 1000x smaller than a finite state machine for even a toy problem), fast to simulate (under 15 seconds for a 559-node network), and can allow a robot in a public space to collect 20 different views of a cup (rotating between each view) and greet 28 visitors in 15 minutes without failure, despite multiple shared resources between the behaviors.

In preparation

Understanding Robot Agency

Agency is the degree to which something or someone can be considered to have free will and autonomy. A robot's appearance of agency depends on both its behavior and the specific person with whom it is interacting. My work in this area contributed to our understanding of how a robot can predict and influence people's perceptions of its agency.

Three different robots are shown; the humanoid upper-torso robot Bandit, a mobile base with a bubble-blower on top of it, and a stand-alone bubble blower.

Agency in Interactions with Children with ASD

In data from children with autism interacting with a mobile bubble-blowing robot, there were two distinct groups of participants: those who treat the robot as an agent, and those who treat it as an object. Bubble-blowing improved speech and the qualitative richness of the interaction only for children with object-like interactions with the robot, whereas the robot's movement increased speech and the quality of the interaction for children with agent-like interactions with the robot.

Published in JHRI


An experimental setup, with Kiwi the robot on a table in the middle of the room, and three chairs arranged around the table, each with a tablet computer in front of it.

Interaction with Families

More agency is not always ``better'': when interacting with families the robot was perceived as more of an agent and as having more emotions when it acted as a competitor in a card game, but interaction between participants was greatest when the robot was in a more passive role such as commenting on a scrapbooking activity.

Published in Ro-Man 2017


The child-like robot Nico is shown, as well as three closeups of Nico's hand showing how he displays the hand symbols for 'rock', 'paper', and 'scissors'.

Effects of Cheating on Agency

Different ways of cheating in a game can change people's perceptions of the robot: a robot that cheated by changing its play was seen as the least honest, but the most agent-like, while a robot that cheated by saying the wrong result was seen as less of an agent but more honest.

Published in HRI 2010


Other Projects

A schematic of the stewart platform design of the SPRITE robot, which is used for the internal components of both Kiwi and Chili.

Design of the SPRITE Robot

The Stewart Platform Robot for Interactive Tabletop Engagement (SPRITE) is a tabletop robot for use in studies with children and families. With 6 degrees of freedom an an expressive phone-based face, this robot was used in a number of the studies described above.

USC CS Department Tech Report



Email is best.

Elaine Schaertl Short
c/o ECE Department
2501 Speedway
EER 6.804, C0806
Austin, TX 78712