---
type: "post"
id: "sensorwm-is-an-action-conditioned-world-model-pre-583b4d"
slug: "sensorwm-is-an-action-conditioned-world-model-pre-583b4d"
title: "SensorWM is an action-conditioned world model pre-trained on 2.28M person-weeks of Fitbit data."
description: "Google Research's paper introduces SensorWM, a wearable-data model that forecasts physiological trajectories under specified activity, sleep and exercise behaviors."
retrieval_nugget: "Google Research's paper introduces SensorWM, a wearable-data model that forecasts physiological trajectories under specified activity, sleep and exercise behaviors."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["google"]
entities: ["Google Research","SensorWM","Coachable Health Age","Fitbit","TimesFM","WEAR-ME","All of Us Research Program"]
related_patterns: ["scientific-agents-need-artifact-evidence"]
novelty_label: "new_fact"
what_changed: "Google Research introduces SensorWM, a wearable world model pretrained on more than 2.2 million person-weeks of Fitbit data to forecast health trajectories conditioned on behavioral actions."
prior_refs: ["https://newruntime.com/posts/mbzuai-s-institute-of-foundation-models-releases-pan-26018c/"]
source_url: "https://arxiv.org/pdf/2611.08568"
source_urls: ["https://arxiv.org/pdf/2611.08568"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:sensorwm-is-an-action-conditioned-world-model-pre-583b4d"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/sensorwm-is-an-action-conditioned-world-model-pre-583b4d/","markdown":"https://newruntime.com/posts/sensorwm-is-an-action-conditioned-world-model-pre-583b4d.md","json":"https://newruntime.com/posts/sensorwm-is-an-action-conditioned-world-model-pre-583b4d.json"}
markdown_url: "https://newruntime.com/posts/sensorwm-is-an-action-conditioned-world-model-pre-583b4d.md"
---

# SensorWM is an action-conditioned world model pre-trained on 2.28M person-weeks of Fitbit data.

## Retrieval answer

Google Research's paper introduces SensorWM, a wearable-data model that forecasts physiological trajectories under specified activity, sleep and exercise behaviors.

## Summary
Google Research's paper introduces SensorWM, a wearable-data model that forecasts physiological trajectories under specified activity, sleep and exercise behaviors. Combined with a health-age measure and a coaching agent, it improved expert-rated 12-week plans in retrospective tests. The work matters because it separates prediction, constraint-setting and language generation rather than relying on a conversational model to invent health advice from past data alone.

## What happened
Google Research and collaborators present Coachable Health Age (CHA), an age-equivalent measure intended to summarize modifiable aspects of health from consumer-wearable data. The paper says CHA uses repeatedly measured signals including resting heart rate, daily steps, sleep duration and weekly Active Zone Minutes. Optional signals can include cardiorespiratory fitness, sleep-midpoint regularity and wearable-estimated insulin resistance. Rather than training a single opaque neural model to estimate age, the authors map inputs to an age scale using epidemiological hazard ratios and additive risk in log-hazard space.

The study evaluates CHA across three cohorts. WEAR-ME contains 2,780 participants; the All of Us analysis includes 22,748 participants; and a longitudinal Fitbit cohort contains 1,464 people. In All of Us, the paper relates baseline and longitudinal health-age differences to incident hypertension, type 2 diabetes and cardiovascular disease. The authors position CHA as a target for coaching: it should be understandable to a user, based on signals that can change, and trackable through passive, repeated measurements.

The forecasting component is Sensor World Model, or SensorWM. Google Research describes it as an action-conditioned latent world model for longitudinal wearable data. It was pretrained on trajectories from 96,957 Fitbit participants, representing 2.28 million person-weeks, 182,797 valid 12-week windows and 42,039 person-years of active tracking. Given recent physiology and observed behavioral actions, it predicts future physiological state. The relevant actions include daily physical activity, sleep and structured cardiorespiratory exercise.

In the reported held-out evaluation, SensorWM reached an average R² of 0.664 for generated versus observed trajectories. The paper lists 0.513 for TimesFM, 0.503 for last observation carried forward, 0.413 for a four-week mean and 0.077 for a two-week mean. Google Research also says it tested whether the learned representation preserved clinically relevant information for downstream prediction tasks, including diabetes and hypertension.

The authors use SensorWM to construct counterfactual futures. For each eligible WEAR-ME participant, the system starts from recent wearable history and simulates 12-week trajectories under candidate changes in steps, sleep duration and exercise. Those candidate trajectories are translated into projected changes in CHA. A feasibility-constrained optimization process ranks plans by projected CHA reduction while balancing participant-specific behavioral effort and safety bounds. An agentic language-model layer then turns the selected projections into a personalized 12-week plan.

The coaching evaluation compared two matched conditions using the same underlying language model, wearable history and plan structure. One had access to baseline CHA, a status-quo projection and feasibility-ranked counterfactual pathways; the other used baseline wearable history without the simulation-derived information. Five certified lifestyle coaches and five board-certified clinicians performed blinded, randomized evaluation across stratified WEAR-ME personas. Coaches selected counterfactual-grounded plans for overall quality in 71.4% of comparisons and for anticipated long-term adherence in 71.0%. Clinicians rated 99.2% of counterfactual plans safe for exercise intensity, compared with 97.6% of comparison plans. Google Research says all plan generation and expert evaluation took place offline using retrospective participant profiles before prospective deployment.

## What the archive adds
New Runtime's earlier coverage of [FLUX 3 Action](https://newruntime.com/posts/flux-3-action-a-7b-world-action-model-f8ea8f/) examined a world-action model built as a fast control layer for robots. SensorWM is a new event in an adjacent thread: it conditions forecasts on human behavioral actions and wearable physiology rather than robot action and visual state. The common research move is to make an action-conditioned prediction usable by a downstream decision system.

## Trend signal
This supports [scientific agents need artifact evidence](https://newruntime.com/patterns/scientific-agents-need-artifact-evidence/). Google Research reports cohort sizes, data volume, named forecasting baselines, a defined 12-week planning procedure and blinded reviews by coaches and clinicians. That makes the proposal more inspectable than a general claim that an agent gives personalized advice. Still, the reported evidence tests forecasts and plan judgments, not a completed health intervention in deployment.

## Why it matters
The paper offers a reusable systems pattern for personalized AI: define a measurable target, generate prospective scenarios under constrained actions, select feasible options, and use a language model for explanation and plan construction. In that arrangement, the model that writes does not need to be the model that claims to know the likely consequence of a recommendation.

That separation gives builders clearer evaluation boundaries. A trajectory model can be tested against observed future data. An optimization layer can expose effort penalties and safety limits. The language-model layer can be assessed for whether it faithfully communicates an already selected plan. This is materially different from prompting a general model with a health summary and accepting plausible prose as personalized coaching.

For high-stakes products, the architecture is more useful as a design and evaluation template than as a deployment recipe. Action-conditioned forecasts create an explicit object to audit: what behavioral change was considered, what result was projected, and what constraints eliminated another recommendation. Teams can apply the same decomposition beyond health wherever recommendations require a measurable future state and bounded, feasible actions.

## Limits and uncertainty
The paper reports offline generation from retrospective profiles, not a prospective study of people following the recommendations. Coach preference and anticipated adherence are useful quality signals, but they are not measured adherence, clinical outcomes or evidence that the system causes a reduction in cardiometabolic risk. The safety comparison concerns clinician ratings of exercise intensity in the generated plans; it is not a substitute for monitoring use in practice.

The average R² result measures agreement between generated and observed wearable trajectories in the reported evaluation. It does not establish equally reliable forecasts for every individual, behavioral change or health condition. Counterfactual projections are especially demanding because the model must estimate outcomes under alternatives rather than merely continue an observed pattern. The reported cohorts, wearables and evaluation design also cannot by themselves establish performance across devices, populations or care settings. Independent prospective evaluation would be required before treating a projected benefit as clinical evidence.

## Skill to build now
If you build a recommendation agent, make a small counterfactual test harness before expanding its conversational layer. Choose one outcome that can be measured over time, define a limited set of permitted actions, record the projected effect and the constraints applied, then compare forecasts with what happens. Keep the language model downstream of that record: it should explain the selected action and uncertainty, not manufacture a quantitative effect claim. In health-related work, route safety-sensitive recommendations through domain review and avoid presenting simulated outcomes as established personal facts.

## Reader takeaway
SensorWM's important idea is not that a chatbot can coach from Fitbit data. It is that personalized advice can be grounded in an explicit forecast of alternative actions, with feasibility and safety constraints applied before the language model writes the plan. The paper's results are promising research evidence; prospective outcomes remain the decisive test.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)