FEAST: A Flexible Mealtime-Assistance System Towards In-the-Wild Personalization

Rajat Kumar Jenamani, Tom Silver, Ben Dodson, S. Y. Tong, Aimin Song, Yuting Yang, Ziang Liu, Bill Howe, Aimee Whitneck, Tapomayukh Bhattacharjee2025 · DOI

TL;DR

FEAST is a modular, LLM-powered assistive feeding system designed to be personalized "in-the-wild" by users with diverse physical disabilities. By combining parameterized behavior trees with large language models (LLMs) for code synthesis and transparency, FEAST allows users to customize robot behavior—like speed, distance, and even custom head gestures—using simple natural language commands.

The Challenge of Personalized Care

While roboticists have made strides in automated feeding, most existing systems are "one-size-fits-all." In reality, mealtime is deeply personal and context-dependent. A user with Multiple Sclerosis has different physical needs than someone with a Spinal Cord Injury; furthermore, the same user might want their robot to behave differently when eating alone while watching TV versus dining at a social gathering.

Current systems often lack the adaptability to handle these nuances, the transparency to explain their actions to non-expert users, and the safety frameworks required for close-proximity human-robot interaction in home environments. FEAST aims to bridge this gap by providing a flexible framework that care recipients can tailor to their specific needs on the fly.

Methods: Adaptability, Transparency, and Safety

The researchers developed FEAST based on three core tenets designed to facilitate real-world deployment.

The three core tenets of FEAST: Adaptability, Transparency, and Safety.
  • Modular Hardware: The system uses a Kinova Gen3 arm with a custom tool-changer. It can swap between a novel feeding utensil (with integrated motors for better orientation), a drinking mug, and a mouth-wiping tool.
  • Parameterized Behavior Trees (BTs): Instead of rigid code, robot skills are represented as BTs with adjustable parameters (e.g., Speed, DistanceToMouth).
  • LLM-Based Personalization: When a user makes a natural language request (e.g., "Feed me faster"), an LLM (GPT-4o) translates that request into structured updates for the BT parameters. It can also synthesize code for new "gestures" (like a specific head shake) to trigger robot actions.
  • Transparency via Summarization: To keep the user "in the loop," the LLM provides natural language summaries of what the robot is doing and how it has been modified, aligned with IEEE Transparency Standards.
  • Safety Architecture: The system includes multi-layered safety checks, including static code verification of LLM outputs, workspace limits, force-torque monitoring for compliant control, and physical emergency stops, all aligned with the ISO/TS 13482 safety standard.

Evaluation and User Diversity

The paper focuses on real-world human-centered data rather than massive pre-existing datasets. A formative study with 21 care recipients across a wide range of conditions helped identify 46 distinct types of personalization requirements.

The formative study included 21 care recipients with diverse conditions such as Cerebral Palsy, Muscular Dystrophy, and Multiple Sclerosis.

Following this, the team conducted a 5-day in-home evaluation involving two "Community Researchers" (care recipients who co-authored the paper) across six different meal contexts, such as social dining and watching TV. Finally, an Occupational Therapist validated the system against 10 realistic clinical scenarios.

Key Results and Insights

  • High Success Rates: In home trials, FEAST achieved an 89.27% success rate for bite acquisition and a 93.07% success rate for bite transfer. Drinking and wiping tasks were even more successful (up to 100%).
  • Low Cognitive Workload: NASA-TLX scores indicated that the system was easy to use, with workload scores significantly lower than established baselines for assistive technology.
  • User Preference: Both community researchers reported that FEAST provided a greater sense of independence and control compared to being fed by a human caregiver.
  • The Power of "In-the-Wild" Personalization: The study revealed that users don't just want a robot that works; they want a robot that fits. For example, one user requested the robot retract further after every bite during a social meal to avoid obstructing their view of their dining partner.

Limitations and Future Work

While successful, the current vision system occasionally struggles with open-set food detection (e.g., mislabeling diverse food items). Future work will focus on more robust fallback mechanisms to prevent LLM hallucinations and developing "long-horizon preference learning" so the robot can remember settings based on the time of day or environment. Additionally, the team aims to improve physical interactions, like skewering slippery food, by incorporating closed-loop tactile feedback.

Summary updated

BibTeX
@misc{jenamani2025feast,
  author = {Rajat Kumar Jenamani and Tom Silver and Ben Dodson and S. Y. Tong and Aimin Song and Yuting Yang and Ziang Liu and Bill Howe and Aimee Whitneck and Tapomayukh Bhattacharjee},
  title = {{FEAST: A Flexible Mealtime-Assistance System Towards In-the-Wild Personalization}},
  year = {2025},
  doi = {10.15607/rss.2025.xxi.083},
}
Figure