1Nanyang Technological University 2Xiaomi EV 3University of Liverpool 4Tongji University 5Tsinghua University 6The Hong Kong Polytechnic University
Corresponding author
The world model literature has grown rapidly since 2023 (a), yet its benchmarks do not yet test the third placement: none of the 42 world model benchmarks and validated learned policy evaluators reviewed scores prediction of a particular robot under interventions chosen for it (b). The survey places a world model inside the robot (I), in place of the world (II) or with an environment side that serves the robot (III), its focus (c).
Fig. 1 of the survey. (a) OpenAlex works in computer science, robotics and driving whose title or abstract contains “world model(s)”, by the document type of their primary location; the line counts matching arXiv cs.* submissions; 2026 runs to October 16, and its type split reflects indexing lag. (b) Bubble area gives the works per half-year of first release (arXiv v1 or publication) that score each criterion; one work can score several; task-stream benchmarks such as LIBERO are not counted. Full-size image
Robots now plan and act with world models, learned predictors of how the world evolves under actions and interventions, and are trained and evaluated inside them. World model related surveys take the robot’s viewpoint, leave each use’s requirements implicit and overlook environments that predict and adapt to the robot; such assistance can be mistaken for robot competence. This survey makes four contributions. First, it places world models in one loop with two decision-makers, where an assignment rule separates agent-centric world models inside the robot, learned worlds in place of the world, and environment-enhanced world models that the environment queries for this robot. Second, it reviews the first two paradigms by query and use: in reported results, usefulness tracks action-faithful, closed-loop prediction more than visual realism, and learned evaluators are validated mainly by rank agreement on a few fixed policies. Third, its coded evidence map of environment-enhanced world models shows that environments predict the robot mainly with outcome critics and preview changes with physics engines and digital twins, rarely with learned world models. Fourth, it proposes a research agenda and an evaluation protocol separating robot learning from environmental assistance, even in learned worlds refit on the robot’s rollouts.
The survey describes embodied systems as one loop with two decision-makers: the robot’s policy chooses actions , and the environment side’s rule chooses interventions , such as what the robot experiences, how its world is configured or what it is told. The same kind of predictor can sit inside the robot, in place of the world, or with an environment side that predicts this robot and adapts to it. Each placement owes its own guarantee.
One loop, three placements (a), and the assignment rule that classifies each use of a world model by three questions asked in order (b).
Fig. 3 of the survey. Uses, not models, are classified; the order is conceptual, not a ranking. Full-size image
Inside the robot
: the joint physical and task state; : the robot’s action. The robot’s own learning algorithm maintains the model.
In place of the world
The conditions are fixed or chosen by humans; the model does not condition on this robot.
With the environment side, serving the robot
: robot variables such as competence; : embodiment; : the environment side’s history (its evidence about the robot).
Both rules are judged by one competence objective, , with the task outcome measured outside the adapting system: robot learning moves , and environment adaptation moves .
Answer the three questions in order; the first match decides. You can also load a system from the taxonomy to see how the survey classifies it.
The 69 leaves of the survey’s taxonomy (Fig. 5 of the survey), grouped by placement and by query, use or prediction question; representative works, not a census. Select a work to see where it sits and how to cite it.
† preprint · ★ strict core · FT fidelity tracking · L2 adaptation ladder level (Paradigm III) · ↔ the same work, model or model family appears in two places
An environment-enhanced world model is a predictive model held by the environment side, a rule other than the robot’s own policy update; in the strict core, this model is a learned world model. It is queried to choose an intervention meant to improve this robot’s learning, safety or task success: what the robot experiences, how its world is configured or what it is told. The term denotes neither data or environment augmentation nor a world model enhanced with environment information.
The survey groups these uses by four prediction questions: Q1 choose experience (what should this robot experience next?), Q2 predict the robot (how will it perform, and where will it fail?), Q3 preview changes (what follows if the environment changes?) and Q4 share predictions (what should the environment tell this robot?).
The evidence map (Fig. 9 of the survey) places 67 coded papers by prediction question and by the type of model the environment side queries. They are the papers that answer one of the four questions among 122 papers on environment adaptation, each coded twice independently and adjudicated (searches of October 12 and 13, 2026): a coded set, not a census. Of the 67, 59 are Paradigm III entries; the other 8, Q1 entries at L0, do not condition on the robot and are reviewed as learned worlds. Only the first column passes the strict world model test.
The survey’s first finding concerns placement: learned world models are rare where the environment side predicts the robot (Q2) or previews changes (Q3). There they account for 2 of 21 and 3 of 18 entries, against 16 surrogates or critics and 11 engines, twins or analytic models. An updated search in which learned world models reach parity with surrogates in Q2, or with engines and twins in Q3, would refute this finding.
Select a cell to list its entries.
On narrow screens the columns are abbreviated: WM world model, Critic surrogate or critic, Beh. behavior model, Eng. engine or twin.
The adaptation ladder grades what the environment side conditions on and learns; fidelity tracking is a flag, not a rung.
Fig. 4 of the survey. Full-size image
On Q1, the line between L0 and L1 separates Paradigms II and III; Q2–Q4 are Paradigm III at every level. A higher level is not intrinsically better: fixed, verified infrastructure may suit settings where continual change would harm reliability.
Follows the filters of the evidence map above.
Eleven representative learned policy evaluators (Fig. 8 of the survey), not a census, placed by the number of policies or checkpoints validated and the reported Pearson r with ground-truth success (for RoboWorld, RoboArena leaderboard scores). Values are taken unimputed from the primary papers; protocols differ, so markers are not a ranking. Validation is mostly rank or correlation agreement over a few fixed policies.
The chart scrolls sideways on narrow screens.
If an environment helps, a rising success rate cannot show that the robot learned. Where the environment side can be frozen or swapped while the robot is held fixed, the survey proposes to cross robot learning, which moves , with environment adaptation, which moves , and to read the interaction. The protocol is a proposed reporting standard, not a validated experiment.
Four checks that separate robot learning from environmental assistance: a 2×2 factorial (a), assistance removal (b), cross-play (c) and a world model update audit (d).
Fig. 11 of the survey. Chips under each panel mark the paradigms a check applies to; (d) applies to I and III only in refit loops. Full-size image
All 304 works the survey cites, checked against their primary pages. Filter by where the survey discusses them, by type or by year, and copy a BibTeX entry.
If you find the survey useful, please cite it as follows.
@misc{li2026worldmodels,
title = {{World Models Across the Embodied Loop: A Survey from
Agent-Centric Prediction to Environment-Enhanced Adaptation}},
author = {Li, Zirui and Ma, Boyu and Liu, Haochen and Jin, Zhehao and
Wang, Zhidong and Yang, Lei and Zhao, Xiaocong and
Shan, Zitong and Liang, Jingsong and Chen, Guang and
Li, Shengbo Eben and Sun, Jian and Yang, Chenguang and
Yang, Jianfei and Liu, Ziwei and Lv, Chen},
year = {2026},
url = {https://lzrbit.github.io/wm-eai/},
note = {Preprint}
}