A beginner reaches for the soy sauce when the recipe wants salt. Most feedback systems tell them afterward. University of Maryland researchers built a wearable AI that acts before the mistake happens, turning the learner’s head or pulling back the arm mid-reach.
The system, called ProtoCall, was announced on November 9, 2026 and is being presented at the ACM Symposium on User Interface Software and Technology (UIST) in Detroit. In a 16-person study, every form of physical feedback beat practicing without it. It is a lab prototype, not a product.
How ProtoCall works
The team, led by PhD student Kyungyeon Lee with Jun Nishida, assistant professor of computer science, built a head- and chest-worn exosuit. Custom eye-tracking glasses and a wrist motion tracker feed a dual-loop vision-language model running on a portable NVIDIA Jetson Thor computer. The system tracks where the user is looking and predicts reaching in real time, so it can spot a likely error before the hand gets there.
Nishida describes the aim as embodied human-AI collaboration. AI has transformed how people work with text, images and code on screens, he said, but many valuable skills, such as cooking, assembling products or operating equipment, are physical.
Two kinds of nudge
ProtoCall uses two physical mechanisms, according to the UMD release.
- Preemptive attentional nudging: if a user stares at the wrong item for more than 800 milliseconds, a lightweight motorized cable on a headband gently turns the head toward the correct object.
- Preemptive reaching nudging: if the user begins reaching for the wrong item, a chest-mounted motor pulls the arm back within 113 milliseconds, stopping the hand in mid-air.
The team compared these with correcting the user only after an error. Nothing here relies on a screen or a voice prompt; the guidance happens through the body itself.
What the study found
Sixteen participants performed a 15-step cooking-themed sequential task. All three forms of physical feedback led to fewer errors than practicing without feedback, both during practice and when participants later did the task on their own. Errors fell progressively from post-error correction to wrist nudges to head nudges, and head nudges produced significantly fewer errors than both of the others. People who received preemptive nudges also reported higher confidence and better awareness of their errors than those corrected after the fact.
The team also found a trade-off. Head nudges gave richer spatial context and helped users build a mental map of the whole sequence, reaching the highest learning accuracy. Wrist nudges intervened faster and preserved the user’s sense of control, but mostly said what not to do. The researchers suggest head nudges for complete beginners in new spatial workflows, then subtler wrist nudges for intermediate users who need fast error prevention without losing pace.
What it means
The most immediate use, Nishida said, is workforce training in industrial manufacturing, where new workers must learn long, complex assembly sequences with few mistakes. Other possible uses named in the release include step-by-step cooking, operating machinery, and helping patients relearn motor functions during physical rehabilitation.
For training teams, the notable idea is timing. Conventional coaching reviews errors after they occur; ProtoCall intervenes in the roughly one-tenth of a second before. For learners, the study suggests that preventing errors, rather than correcting them, built confidence. For product builders, the choice of head versus wrist feedback is a design lever tied to skill level.
ProtoCall extends the team’s earlier FIXical I/O, presented at the ACM CHI conference earlier this year, which used a hand exoskeleton to sense finger errors and correct them while people learned key sequences. ProtoCall carries the same idea to the arm and head and to multi-step physical tasks. At UIST this week, the team is also showing a live demonstration that combines both systems.
- Evidence so far: 16 participants, one cooking-themed task, in a lab
- Open questions: factory settings, comfort over long shifts, and different task types
The bottom line
ProtoCall treats AI as a coach for the body, not a replacement for the worker. As Nishida put it, the same technologies making machines more capable can also help people grow. The early data is small, but the idea of stopping a mistake before it happens is a notable shift for skill training.