CMU, Cornell Develop System To Help Robots Adapt To Natural Flow of Eating

04/15/2026    Mallory Lindahl

The Breakdown :

  • An algorithm driven by wearable sensors advances assistive feeding by predicting when someone is ready for their next bite.
  • It helps robots adapt to real-world eating, including conversation and pauses. 
  • Adjustable settings let users set the pace while easing workload.

* * *

New work led by Carnegie Mellon University researchers could predict when someone is ready for their next bite, bringing assistive feeding devices closer to the rhythms of everyday life for the millions of people who require help with eating.

Researchers in the School of Computer Science and Cornell University developed a new assistive feeding algorithm, Wearable Approach for Feeding With Learned Bite Timing (WAFFLE), that can be integrated with robotic feeding devices to predict when someone is ready for their next bite by interpreting natural behavioral cues.

“WAFFLE represents a shift toward more human-centered assistive robotics. We need systems that do not just perform tasks, but do so in ways that respect the varied rhythms and nuances of everyday life,” said Jessie Yuan, an SCS undergraduate student on the research team. “By centering natural user signals and offering adaptable control, WAFFLE moves closer to assistive feeding that feels less like operating a machine and more like being understood.” 

In the United States, approximately 3.5 million people with motor impairments require help with feeding — a number that will only rise as the general population ages. Assistive robots can help, but determining exactly when to offer the next bite remains challenging.

Bite timing is deceptively complex. Factors like the type of food, how long it takes to chew, whether someone is speaking and the broader context of the meal — such as eating alone versus in conversation — can all dramatically shift timing. Even attentive human caregivers can struggle to consistently interpret these signals, especially in distracting environments.

WAFFLE works by equipping participants with a set of lightweight sensors embedded in glasses and earbuds, along with a commercially available throat microphone. These sensors capture signals associated with everyday eating behaviors such as head movement, chewing, speaking, and vibrations from swallowing and vocal activity. By combining motion data from the glasses and earbuds with vibration data from the throat microphone, WAFFLE can analyze how people naturally engage with food and conversation.

A woman in a pink shirt and glasses sits in her wheelchair at a dining table

When using WAFFLE, participants reported less workload without losing their feeling of control over the system.

To train their system, the team first conducted a controlled study in the Robotic Caregiving and Human Interaction Lab, led by Robotics Institute (RI) Assistant Professor Zackory Erickson. They collected data from 14 participants without motor impairments using the Obi feeding robot. The participants dined alone and in a social situation where they talked while eating. This distinction proved important. In social scenarios, people pause more often, speak between bites and move with less predictable rhythms. WAFFLE uses this varied data to help the robot adapt to different situations.

Once trained, the WAFFLE model produces a continuous prediction of how long it will be until the user is ready for another bite. The predictions are then paired with a user-adjustable assertiveness threshold, which allows individuals to control how proactively the robot offers food, whether they prefer a more cautious pace or a faster, more responsive experience.

The researchers also tested WAFFLE in the real world. In a separate study led by the Cornell team, two participants with motor impairments used the system in their own homes with a Kinova robotic arm. WAFFLE performed effectively with both participants despite differences in environment, feeding rhythm and user preference. 

“This work is especially exciting because it shows that wearable sensors can successfully capture implicit cues from users during feeding, allowing a robot to effectively time the delivery of a bite.” said Akhil Padmanabha, a RI Ph.D. graduate and part of the WAFFLE team. “In addition to showing generalization to a variety of scenarios, our participants using WAFFLE reported less workload without losing their feeling of control over the system.”

Along with Yuan, Padmanabha, and Erickson, the WAFFLE team included recent SCS graduate Tanisha Mehta; Rajat Kumar Jenamani, a Ph.D. student at Cornell; Eric Hu, an undergraduate at Cornell; Victoria de León, a Tecnológico de Monterrey undergraduate and former CMU Robotics Institute Summer Scholar; Anthony Wertz, a CMU Robotics Ph.D. alumni; Janavi Gupta, an SCS undergraduate; Ben Dodson, a Cornell engineering alumni; Yunting Yan, a Cornell Ph.D. student; Carmel Majidi, the Clarence H. Adamson Professor in CMU’s Department of Mechanical Engineering; and Tapomayukh Bhattacharjee, assistant professor in Cornell’s Department of Computer Science.

WAFFLE was accepted to the 2026 ACM/IEEE Conference on Human Robot Interaction, where it won a best paper award. To learn more about the work, visit the project website.

For More Information: Aaron Aupperlee | 412-268-9068 | aaupperlee@cmu.edu