Source-linked AI summary
OnBodyAI - Why the Next AI Needs a Body
Michael Beigl
TL;DR
Wearable sensors can capture personal body data that current assistants still ask users to provide, but sensing and interpretation have limits. This article argues for on-body AI that couples sensing, interpretation, and actuation under the wearer’s control, and identifies hardware, efficient models, and open platforms as needed progress. Its value depends on reliable everyday signals, validation on unseen people, and auditable systems that can produce lasting individual benefit.
Problem
Method
The article sets out an OnBodyAI research agenda connecting body sensing, interpretation, and actuation under the wearer’s control, alongside progress in hardware, efficient models, and open platforms.
Source: §11 · pp. 9–10Results
The article concludes that on-body AI’s value depends less on model size than on coupling sensing, interpretation, and actuation under the wearer’s control.
Source: §11 · p. 10Takeaways & Limitations
Lasting individual benefit requires reliable everyday physiological sensing, models validated on unseen people, and open platforms that make results reproducible and claims auditable.
Source: §11 · p. 10
Abstract
from arXiv · showhide
The next AI will not win by knowing more about the world. It will win because it knows you: because it lives on your body, works for you and answers to you. That is what the term OnBodyAI stands for. It describes what happens when large language models are given a body, and what research and industry need to make that happen. This article is based on my talk in the Industrial Panel at UbiComp/ISWC 2026 in Shanghai (October 13, 2026) and follows its ten key statements.
1 AI HAS A BRAIN, BUT NO BODY
Wearable data can already inform language-model advice, but assistants still ask users for information that nearby body sensors can measure. The author argues that physical-context personalization is an industry differentiator that remains weakly occupied.
1 AI HAS A BRAIN, BUT NO BODY
- Health-LLM, PH-LLM, SensorLM, and an Apple watch foundation model illustrate wearable-data inputs for language-model prediction, coaching, or sensor-signal interpretation.The passage describes the pattern as sensor data in, advice out.Source: §1 · p. 1
- Current assistants ask how users slept or feel even when nearby sensors can measure heart rate, sleep, jaw clenching, and voice tension.The author summarizes the gap as AI having a brain but no body.Source: §1 · p. 1
- Physical-context personalization is the next AI differentiator for industry, while body data remains a weakly occupied gap.Source: §1 · p. 2
2 THE OTHER HALF: WE HAD THE WEARABLE, WE LACKED THE BRAIN
Wearables long had sensing and context-awareness, but their interpretation relied on hand-written rules; learned on-device interpretation now pairs with that body-sensing base. For products, intelligence should fit familiar form factors and habits rather than ask people to change.
2 THE OTHER HALF: WE HAD THE WEARABLE, WE LACKED THE BRAIN
- OpenEarable 2.0 carries seven ear sensors with learned on-device interpretation, pairing the community’s wearable expertise with AI.The MediaCup already had sensors, radio, and battery in 1999, but hand-written rules prevented it from learning anything new.Source: §2 · p. 2
- The author describes 25-year-old ideas as finally shippable, opening a product window for on-body AI.Source: §2 · p. 2
3 GREAT POTENTIAL IS NOT YET A SUCCESSFUL PRODUCT
- Wearable product ideas can fail despite technological promise, as Google Glass, Microsoft Band, and Humane AI Pin illustrate.Each asked people to change; the passage contrasts them with earbuds, rings, and watches that add intelligence to familiar objects and habits.Source: §3 · pp. 2–3
- The industry perspective is that intelligence belongs in form factors people already wear, avoiding a new device or habit.Source: §3 · p. 3
4 THE EAR IS THE NEW WRIST
The ear combines a large installed base of familiar devices with access to diverse body signals. The author presents it as a promising sensing site, while noting that consumer earbuds must sustain clinically capable signals throughout the day.
4 THE EAR IS THE NEW WRIST
- About 302 million hearables out of roughly 504 million wearables shipped in 2023—about 60 percent—represent a substantial installed base for ear sensing.The passage also notes Apple added heart-rate sensing and live translation to AirPods Pro 3 in September 2025.Source: §4 · p. 3
- EarXplore publications on ear-based gestures grow from year to year.Source: §4 · p. 3
- More than 30 phenomena can be sensed at the ear, including pulse, oxygen saturation, blood pressure, voice, chewing, muscle tension, and electrical signals from eyes and brain.The author highlights the ear’s proximity to vessels, jaw, eyes, and brain, and its stable position during extended earbud wear.Source: §4 · p. 3
- The industry opportunity is health sensing through consumer earbuds, provided clinically capable signals hold up all day.Source: §4 · p. 3
5 BEING ON THE BODY IS NOT THE SAME AS UNDERSTANDING THE BODY
Wearable measurements can look acceptable by some metrics while failing to generalize to unseen people. The author argues that sensor quality and validation across people and settings constrain what body-sensing AI can understand.
5 BEING ON THE BODY IS NOT THE SAME AS UNDERSTANDING THE BODY
- R² ≈ 0 on unseen people for blood-glucose models from PPG, despite more than 90 percent of estimates in Clarke zones A and B even for the mean predictor.The passage says models that look good on randomly mixed data collapse on people they have not seen, and concludes that current metrics hide the failure.Source: §5 · p. 4
- Six commercial wrist wearables achieve Cohen’s kappa of only 0.21 to 0.53 against polysomnography; the passage characterizes this as fair to moderate agreement, below what medical diagnoses require.The same passage reports mean absolute percentage error above 30 percent for every wearable brand in a 65-study energy-expenditure review.Source: §5 · p. 4
- A larger model cannot reconstruct uncaptured information; contact, motion, and signal quality constrain what it can know.Models must be validated on unseen people, situations, and recording environments, not only random splits.Source: §5 · p. 4
6 CLOSING THE LOOP: AI GIVES YOU FREEDOM
On-body closed loops can sense the user and act on the body under the user’s rules, potentially reducing cognitive effort. Examples span stress-responsive meditation, heart-rate-triggered warming, and heatables aimed at comfort.
6 CLOSING THE LOOP: AI GIVES YOU FREEDOM
- The closed loop should act on the user’s terms and rules, relieving cognitive effort rather than taking control away.The passage frames user control and reduced cognitive load as the loop’s intended value.Source: §6 · p. 4
- Calmables measures heart rate at a ring and warms the ear with infrared earables when the rate indicates a need to relax.Its sensing and action occur at different body sites: finger and ear.Source: §6 · p. 4
- EarStreAM uses heart rate and heart-rate variability to detect rising stress, then adapts an LLM-generated meditation in real time to the body’s response.Users’ trust is decisive to whether they hand over control.Source: §6 · p. 5
- Heatables participants felt about 1.5 °C warmer, while the industry case is that output channels must deliver lasting user value.Fig. 6 identifies Heatables experimental conditions and a prototype; Fig. 5 depicts the Calmables and EarStreAM demo posters.Source: §6 · p. 5, Figure 6 · p. 5, Figure 5 · p. 5
7 ON THE BODY, MILLIWATTS MATTER MORE THAN MODEL PARAMETERS
On-body AI faces a stringent energy budget: an earbud must run all day while dissipating heat near skin. The paper argues for processor-aware model design rather than relying only on generic model shrinking.
7 ON THE BODY, MILLIWATTS MATTER MORE THAN MODEL PARAMETERS
- About 30 milliwatts is the average whole-earbud budget for a working day, versus around 700 watts for a data-center GPU—a four-orders-of-magnitude gap.Fig. 7 compares an NVIDIA H100 SXM TDP with an earbud battery of about 50 mAh over about six hours.Source: §7 · p. 6, Figure 7 · p. 6
- Each added milliwatt brings heat to dissipate near skin and consumes scarce on-body volume.The passage uses the ear as an example of this volume constraint.Source: §7 · p. 6
- Generic quantization and pruning are not optimal when embedded processors’ runtime and energy properties are not considered; MicroNAS is a first step toward target-processor-specific model design.The passage links this design need to battery life as a demanding wearable-product requirement.Source: §7 · p. 6
8 AI EATS SOFTWARE. WEARABLES ARE STILL BUILT BY HAND
Wearable development remains a hands-on chain of hardware work and troubleshooting, unlike increasingly AI-assisted software production. The proposed response is to automate part of the embedded stack and share hardware.
8 AI EATS SOFTWARE. WEARABLES ARE STILL BUILT BY HAND
- Wearable hardware still requires hands-on assembly and repeated troubleshooting when prototypes diverge from simulations or parts and layouts change.The passage gives fine-pitch soldering, unavailable components, physical interference, and unexpected line noise as examples.Source: §8 · p. 6
- The lab’s intermediate prototype state illustrates that building wearables remains a physical process.Fig. 8 is captioned as a typical intermediate prototype state.Source: Figure 8 · p. 7
- The proposed response is to automate part of the embedded stack, use browser-based tooling from recording to a wearable-running model, and share hardware.The passage says embedded engineering labor is scarce and identifies automation and open hardware as levers for cost and time-to-market.Source: §8 · p. 7
9 SIGNALS BECOME MEANING – BUT ONLY IF THE HARDWARE PROVIDES THE BASIS
Learned models can interpret weak, noisy, personal body signals, but only when the hardware captures and exposes them. The examples demonstrate ear-based sensing while also showing that closed devices limit what can be learned.
9 SIGNALS BECOME MEANING – BUT ONLY IF THE HARDWARE PROVIDES THE BASIS
- EarRumble classification reaches 95 percent; earEOG tracks horizontal but not vertical eye movement, and OpenEarable ExG captures EEG, EMG, and EOG.The section presents these as examples of learned models applied to weak, noisy, personal signals.Source: §9 · p. 7
- The model can only learn what it receives: a closed earbud provides steps and heart rate, and none of the three examples would have been possible on that closed device.Fig. 9 is captioned “From signal to meaning” and is based on EarRumble, earEOG and OpenEarable ExG.Source: §9 · p. 8, Figure 9 · p. 8
- None of the three examples would have been possible on a closed device, and closed interfaces limit what AI can learn and the system’s impact.The passage characterizes raw sensor access as a product decision that creates a balancing act for industry.Source: §9 · p. 8
10 CLOSED SYSTEMS GO TO ONE MARKET, OPEN SYSTEMS GO EVERYWHERE
Open systems support trust and wider reuse, while open platforms can also move from community projects into regulated products. The OpenEarable family and other examples illustrate this path across devices, labs, and applications.
10 CLOSED SYSTEMS GO TO ONE MARKET, OPEN SYSTEMS GO EVERYWHERE
- Openness supports reproducibility, auditable privacy claims, and faster product development, while enabling use beyond a single vendor’s market.The passage links openness to trust for devices worn all day that listen to heart and voice.Source: §10 · p. 8
- The OpenEarable family spans ear sensing, ExG signals, and brain-computer interfaces, with OpenEarable 2.0 offered as an open-source hardware and software platform.Fig. 10 shows the 2.0, ExG, and BCI family.Source: §10 · p. 8, Figure 10 · p. 9
- Open devices and products include OpenBCI, Omi, and PineTime; the community has held an OpenWearables workshop since 2024.Other examples named include the -Ring and BioGAP-Ultra, whose design files are open.Source: §10 · p. 9
10.1 Open platforms are the on-ramp to real markets
- Loop, Nightscout, and Tympan illustrate open platforms reaching regulated care, community use, and hearing applications; reference platforms can cut prototyping time and cost.Tidepool Loop received FDA clearance in January 2023.Source: §10.1 · p. 9
11 CONCLUSION: GIVING AI A BODY, AND GIVING PEOPLE THE CONTROLS
OnBodyAI’s value depends less on model size than on how sensing, interpretation, and actuation are coupled and controlled by the wearer. The paper calls for progress in sensing, efficient and generalizable models, and open platforms to deliver lasting individual benefit.
11 CONCLUSION: GIVING AI A BODY, AND GIVING PEOPLE THE CONTROLS
- The envisioned system lives on the body, works for the wearer through a closed loop under their rules, and is open, low-power, and candid about sensing limits.The ear is identified as a place people already wear devices.Source: §11 · p. 9
- OnBodyAI’s central claim is that value depends less on model size than on coupling sensing, interpretation, and actuation under the wearer’s control.The paper presents OnBodyAI as a research agenda rather than a product category.Source: §11 · pp. 9–10
- Realizing the agenda requires reliable everyday sensing, milliwatt-budget models validated on unseen people, and open platforms for reproducibility and auditability.Only when these fronts come together, the paper argues, can on-body AI move from data collection to measurable, lasting individual benefit.Source: §11 · p. 10