A new system by Yun Ho, Romain Nith, and Pedro Lopes combines AI and electrical muscle stimulation to physically guide users through unfamiliar tasks. This represents a leap forward to general purpose, context-aware physical assistance.
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 muscles gently guiding your movements, as if an invisible teacher was there lending you his know-how. Imagine that same feeling helping you twist open a child’s medicine bottle, operate a camera, or perform tasks you’ve never practiced before.
This is not science fiction. This is the vision realized by doctoral students Yun Ho and Roman Nis under the supervision of Associate Professor Pedro López of the University of Chicago’s Department of Computer Science. Their work, which recently won the Best Paper Award at the ACM CHI 2026 conference, is gaining attention across the human-computer interaction community.
From niche gadgets to general-purpose physical support
Electrical muscle stimulation (EMS) is not new. For years, researchers have strapped electrodes to the body and delivered controlled electrical currents to teach piano sequences, demonstrate sign language, and support stroke rehabilitation. What about the prey? These systems are always like training wheels, useful only for the limited tasks for which they are programmed, unable to instantly adapt to the messy and unpredictable real world.
As the research team notes, traditional EMS support has been “highly specialized, static, and context-independent.” In other words, the “instructions” of the muscles only fit into the situation that the designer anticipated. Ask EMS to help you shake the can of spray paint, and the spray paint will start moving. Even if you hold up a spray can of cooking oil, this device is clueless. Because I don’t need to shake it and I don’t understand why.
But this new system, which the authors have dubbed “embodied AI,” signals a change. By harnessing the power of modern multimodal artificial intelligence (think computer vision with vision models like CLIP and GPT-4 level reasoning), it integrates what you’re looking at, where you are, and even your body pose to generate movement guidance tailored to the moment. EMS no longer follows recipes. It improvises with you.
I’m interested in how people understand and develop relationships with devices that communicate through bodily movement (as opposed to audio or visuals). “Embodied AI” allowed me to explore this issue in the area of physical assistance. It was particularly insightful that participants were able to “think aloud” while using our system and learn how to interpret the movements caused by the machine. ”
Yoon Ho, PhD student, Department of Computer Science, University of Chicago
AI that not only “knows that” but also “knows how”
The magic here is procedural knowledge. That is, a hard-to-describe sense that is embodied in how to do something, such as squeezing the lid of a jar just right and twisting it open, or unlatching a European window using a combination of wrist and shoulder movements. For decades, researchers have focused on giving people factual information. This is an approach that directly conveys “know-how” to the muscles.
What will change with context-aware generative EMS?For the first time, users will be physically guided through unfamiliar and complex physical tasks, even if they can’t explain what they need. The paper details a user study in which participants successfully opened a medicine bottle with a locking mechanism, took a photo with a disposable camera, and used an avocado tool guided by dynamically generated muscle cues. And if the AI made a mistake (deliberately for testing purposes), the human would notice, adapt, and come up with a solution by re-prompting the system or correcting the error.
This iterative approach, where bodily intuition meets AI suggestions, is key. One participant said, “Body intuition helps you notice mistakes quickly,” giving you an advantage over reading step-by-step instructions or watching videos.
“This could be a game-changer not only for highly physical tasks (such as learning the physical skills needed for manufacturing or handling materials, learning to play a musical instrument), but also in situations where the user may be situationally impaired (such as multitasking, performing multiple gestures at the same time, or being unable to see in the dark),” Lopez said.
Who will benefit? Everyday applications of embodied AI
The vision for this technology goes far beyond laboratory demonstrations. Consider a sector where procedural know-how is essential and mistakes are costly.
- Healthcare and rehabilitation: Imagine a physical therapist’s patient practicing unfamiliar movements at home or an elderly person using assistive devices. Instead of following a booklet or video, you can coach your muscles through safe biomechanics.
- Industrial and skilled workers: Workers retooling for new equipment are physically guided through muscle movements, reducing injury risk and training time.
- Accessibility: While blind and low-vision users already benefit from AI systems that describe scenes, context-aware EMS could provide direct physical guidance to transform ordinary environments into accessible ones by teaching them new gestures and tasks.
- Daily life: From travelers grappling with “foreign” appliances to hobbyists assembling unfamiliar gadgets, this system provides flexible support tailored to the field and task at hand.
Lopez and the researchers spoke candidly about current limitations, including electrode calibration, the tingling sensation of EMS, and the fact that finesse and “muscle memory” cannot always be solidified with stimulation alone. But as the team points out, advances in AI and EMS hardware are rapid. The trend is clear. In the not-too-distant future, AI guides worn on the body could become as common as wearable health trackers.
“We’re really excited about our system, but this is obviously just the first step and more needs to happen,” Lopez said. “This includes many needed improvements in AI inference, multi-model models that can not only see but feel what the user is feeling (e.g., understand tactile and motion perception). It also includes years of work to make EMS more comfortable, easier to wear, and easier to adjust. Now, this is not just something you can wear in everyday life, but a superhero suit that researchers are experimenting with in the lab.”
Embodied AI Platform Recognized at CHI 2026
We would like to remind our readers that this system does not replace audiovisual guides, but rather complements the human learning that enriches them and the interaction with technology that works through our own bodies. The research will open source the code and invite others to extend, critique, and iterate on the model.
While the field grappled with ethical questions such as who controls the guidance and when misguidance could pose a risk, the team focused on user control and safety by ensuring the AI only works when invited and allows participants to interrupt or adjust at any stage.
Recognized as the best paper at the ACM CHI 2026 conference, this work proves where AI-powered interactions are heading, from passive commands and “smart” environments to true embodied co-pilots performing everyday life tasks.
