Imagine traveling in a foreign country, reaching for a window you’ve never seen before, and instead of struggling to open it, you feel your own muscles gently guide you through the motion, as if an invisible teacher was there, lending their know-how. Now picture that same sensation helping you twist open a child-proof pill bottle, operate a camera, or perform tasks you’ve never practiced before.
This is not science fiction. It’s the vision realized by PhD students Yun Ho and Romain Nith, under the supervision of associate professor Pedro Lopes in the Department of Computer Science at the University of Chicago. Their work, recently honored with the Best Paper Award at the ACM CHI 2026 conference, is turning heads across the human-computer interaction community.
Electrical muscle stimulation (EMS) isn’t new; for years, researchers have been strapping electrodes to bodies, delivering controlled currents to teach piano sequences, demonstrate sign language, or support stroke rehabilitation. The catch? These systems have always been more like training wheels, only useful for a narrow set of programmed tasks, incapable of adapting on the fly to the messy, unpredictable real world.
As the research team puts it, EMS assistance to date has been “highly-specialized… fixed, and non-contextual.” In other words, your muscle “instructions” only fit the situations a designer anticipated. Ask the EMS to help shake a can of spray paint, and it springs to life. Hold up a spray can of cooking oil, and the device is clueless—because it can’t understand that you don’t need to shake it, and why.
This new system, which the authors nickname as “embodied AI”, however, marks a shift. By tapping into the power of modern multimodal artificial intelligence (think vision models like CLIP and GPT-4-level reasoning with computer vision) it merges what you see, where you are, and even your body’s pose to generate movement guidance tailored to the moment. EMS no longer follows a recipe; it improvises alongside you.
“I am curious about how people understand and build relationships with devices that communicate with them through body movements (rather than audio/visual),” said Ho. “In ‘embodied AI’, I got to explore this question in the realm of physical assistance. It was especially insightful to have participants “think aloud” as they used our system and learn how they interpret machine-induced movements.”
The magic here is procedural knowledge—the embodied, hard-to-describe sense of how to do something: gripping a jar lid just right to twist it open, or combining wrist and shoulder movements to unlatch a European window. For decades, researchers focused on giving people factual information; this approach transmits “know-how” to the muscles directly.
What changes with context-aware, generative EMS? For the first time, users get physically guided through unfamiliar, complex physical tasks, even when they aren’t able to explain what they need. The paper recounts a user study where participants succeeded at opening pill bottles with locking mechanisms, snapping pictures with a disposable camera, or using an avocado tool guided by dynamically generated muscle cues. And in cases where the AI made mistakes (on purpose, for the sake of testing) people noticed, adapted, and figured out solutions by re-prompting the system or correcting for its errors.
This iterative, back-and-forth approach where the body’s intuition and the AI’s proposals meet is significant. As one participant described: “the body’s intuitions help notice errors right away,” offering an edge over reading step-by-step instructions or watching a video.
“This could be a game-changer, not only for tasks that are highly physical (such as learning physical skills required for working with manufacturing and materials, or learning musical instruments) but also in situations where users might be situationally impaired (e.g., multitasking and performing several gestures at once, or cannot see in the dark, and so forth),” said Lopes.
The vision for this technology stretches well beyond laboratory demos. Consider sectors where procedural know-how is vital and mistakes are costly:
Lopes and the researchers are candid about present limitations: electrode calibration, the tingling sensation of EMS, and the fact that finesse and “muscle memory” can’t always be hardwired by stimulation alone. But as the team notes, progress in AI and EMS hardware is rapid. The trend is clear: in the not-so-distant future, on-body AI guides may be as common as wearable health trackers.
“While we are really excited about our system it is clearly just the first step, much more needs to happen,” said Lopes. “This includes many needed improvements on AI reasoning, on multimodel models that not only can see but also feel what users feel (e.g., have an understanding of tactile and movement perception), but also the long standing of making EMS more comfortable and easy to wear and calibrate—currently this is not something you can just wear on your everyday life, but more a superhero suit that researchers are experimenting in at the lab.”
It’s worth reminding readers: this system doesn’t seek to replace audiovisual guidance, but to complement it—enriching human learning and interaction with technology that works through the medium of our own bodies. The research open-sources their code, inviting others to extend, critique, and iterate on their model.
As the field grapples with ethical questions, like who controls the guidance and when might misguided instructions pose risks, the team has emphasized user control and safety, with the AI only acting when invited, and participants able to interrupt or adjust at every step.
This work, recognized as the Best Paper at the ACM CHI 2026 conference, stands as a testament to where AI-powered interaction is headed: from passive instructions and “smart” environments, to truly embodied co-pilots for the tasks of daily life.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。