Embedded Files
Maximilian Du*, Jennifer Grannen*, Shuran Song, Dorsa Sadigh
ICRA 2026 | Paper Link
When we want to improve our fitness or athletic skills, we might turn to a personal trainer. Their physical presence allows them to guide us through exercises, provide intuitive feedback, and even give encouragement when the going gets hard.
Robots are also physically embodied agents, capable of many athletic feats. Can they also help with athletic training?
We look at a case study of interval training on a race track. Maintaining a reliable pace is challenging for novice runners without guidance. We modified a robot quadruped to guide runners through this exercise.
In our experiments, we compared user performance between the quadruped and common wearable technology. We find that users' speeds are more consistent and accurate while using the robot. Users also perceived the robot more positively across many subjective axes, including helpfulness, fun, and ease of use.
"Snoopie" the pacerbot is a modified Unitree Go2 quadruped that has a 360-degree camera capable of looking forward to follow lines on a racing track while looking backward to keep an eye on the user.
The Go2 robot controller is capable of running at paces up to 2.6 m/s, a speed sufficiently challenging for most novice runners looking to work on endurance.
For safety and transparency, a back-facing screen will play tones before the robot changes speeds. The screen also displays information related to the robot's state, so all motion is predictable. The screen and the camera are mounted on mechanical fuses that disconnect if excessive force is applied.
Snoopie needs to know their physical abilities to create an exercise that is challenging yet feasible for each user. To accomplish this, we split the user interaction into two phases: speed calibration and interval training.
During speed calibration, Snoopie uses the back-facing camera to track a code worn by the user. The robot speeds up and slows down to maintain a constant follow distance. At the end, the average speed is the user's baseline speed.
During interval training, the robot alternates between a challenge (fast) and recovery (slow) pace calculated from the baseline speed. Users are instructed to follow the robot to the best of their ability.
We conducted experiments under two settings: 1) an N=10 user study comparing performance of Snoopie against wearable technologies and an unguided control condition, and 2) an extended interaction study consisting of one participant over multiple, long sessions. For both settings, we track user speeds as they run the interval exercise. We are looking for speed consistency (variance) and accuracy (error from target). At the end of each session, users fill out a qualitative survey about their experience with each tested technology.
Users were able to run more accurately (A) and consistently (B) compared to matching paces with wearable technologies or the technology-free baseline. Notably, by following the embodied robot, their speed transitions between intervals are sharper (C).
After using the robot, users reported preferences over wearable technology across all tested axes, including ease of operation, intuition, and trustworthiness. For quotes from the user's written feedback, refer to our paper.
As part of his surgery rehabilitation process, a participant ran a total of 11.0 kilometers across 5 sessions with Snoopie.
Across this extended interaction, the participant's positive perception of Snoopie remained mostly consistent, showing initial signs that Snoopie has potential as a long-term training tool.
In his written feedback, the participant noted that the robot was helping make the running process feel less burdensome and even providing some feeling of companionship. For quotes, refer to our paper.
Snoopie and the interval running case study surfaced two key findings for human-robot interactions for physical training.
Calibration to Customize: To create customized exercises that are challenging but feasible for the user, the user must demonstrate their physical abilities. With a physical embodiment, this becomes much easier, like the speed calibration phase of Snoopie's exercise.
Physically Embodied Companionship: It doesn't take highly complex interactions to create a feeling of companionship. The physical presence, combined with some adaptability and visual appeal, can be enough to see a robot as a training partner that both demonstrates and motivates.
@inproceedings{du_grannen2026snoopie,
title = {Will People Enjoy a Robot Trainer? A Case Study with Snoopie the Pacerbot},
author = {Du, Maximilian and Grannen, Jennifer and Song, Shuran and Sadigh, Dorsa},
booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
year = {2026}
}
Google Sites
Report abuse