Logo image
Back
Process-aware robotic food assembly: Closed-loop visual monitoring across heterogeneous ingredients
Journal article   Open access   Peer reviewed

Process-aware robotic food assembly: Closed-loop visual monitoring across heterogeneous ingredients

Ran Yang, Zijian An, Lifeng Zhou and Yiming Feng
Journal of agriculture and food research, v 31, 103286
Oct 2026
DOI: https://doi.org/10.1016/j.jafr.2026.103286
Featured in Collection :   Drexel's Newest Publications
url
https://doi.org/10.1016/j.jafr.2026.103286View
Published, Version of Record (VoR) Open

Abstract

Embodied AI Food service automation Process integrity Robotics
In this study, process cognition refers to the ability to monitor the state of the current food-placement subtask and autonomously advance through a predefined assembly sequence without manual intervention during execution. Current robotic systems address isolated operations but cannot navigate sequential procedures where each step depends on successful completion of the previous one. Here, we propose and evaluate a process-aware vision-language-action (VLA) model built on π0 that combines step-specific language instructions, a predefined sequential controller, and learned visual monitoring for automatic step transitions. In a system-level comparison on a seven-step salad-packing task, the process-aware VLA completed 10 of 10 trials, whereas the unaugmented end-to-end π0 baseline completed 3 of 10 trials. Run-level binary predictive entropy varied across the seven subtasks, indicating task-dependent differences in monitoring uncertainty, while associations with individual food properties were primarily descriptive. No sequence errors were observed in the ten process-aware VLA trials, indicating improved procedural consistency within the evaluated setting, while the generation of timestamped step-level records may support future HACCP-aligned process monitoring. These results provide a basis for developing process-aware VLA systems with autonomous sequence execution and step-level traceability. [Display omitted] •Process-aware VLA autonomously executes a predefined food-assembly sequence.•No sequence errors were observed in ten process-aware salad-packing trials.•The integrated system completed 10/10 trials versus 3/10 for the baseline.•Joint training produced lower predictive entropy across seven subtasks.•Step-level records may support future HACCP-aligned process monitoring.

Metrics

1 Record Views

Details

Logo image