1. What embodied AI is (the short version)

• "World model" has three meanings: (1) an environment model for control that predicts how the environment changes with the agent's actions, (2) a model that generates video and so "simulates" a world, and (3) JEPA-type models that learn representations by predicting video features. The same phrase differs in what is predicted and how it is checkedSourced [Sources 1, 7, 6]
• Robot foundation models (VLA) are built by making a vision-language model output "actions": they take images and a language instruction and output robot actions. Training data is a mixture of teleoperation records, human video and simulationSourced [Sources 12, 15, 16, 18]
• "Success in simulation" is not "success on the real robot": physical quantities such as friction, contact, vision and latency all drift. A robot's success rate has a wide range when trials are few, and what "90%" means depends on the number of trialsSourced / Our calculation

2. What a world model is — three uses

"World model" refers to the idea that an AI holds internally how the world changes (including as a result of its own actions). In research, however, there are three different uses. Below they are sorted according to the papers and technical reports.

(1) An environment model for control: from World Models to Dreamer

"World Models" by Ha and Schmidhuber (2018) builds generative models of reinforcement-learning environments that learn a compressed spatial and temporal representation in an unsupervised way. Features taken from it are fed to a very small policy that solves the task, and the authors report that they could even train the agent entirely inside its own "hallucinated dream" generated by the world model and transfer the policy back to the real environmentSourced [Source 1]. Dreamer (Hafner et al.) learns behaviors from trajectories "imagined" in the latent space of a learned world model and reports exceeding existing approaches in data efficiency, computation time and final performance on 20 visual control tasks [Source 2]. Its successor DreamerV3 handles over 150 diverse tasks with a single configuration and is reported as the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula (Nature 640, 2025)Sourced [Source 3].

(2) A "world simulator" made by video generation

There is a view that treats video generation models as simulators of the world. OpenAI's technical report on Sora (15 February 2024) suggests that scaling video generation is a promising path toward general-purpose simulators of the physical world, while the same report lists limits: it does not accurately model the physics of many basic interactions such as glass shattering, and interactions like eating food do not always yield correct changes in object stateSourced [Source 10].

Genie (Bruce et al., 2024) is a "generative interactive environment" trained in an unsupervised manner from unlabelled Internet videos; at 11 billion parameters it is positioned as a foundation world model, made of a spatiotemporal video tokenizer, an autoregressive dynamics model and a latent action model. The paper says users can act frame by frame in the generated environments although training used no ground-truth action labels [Source 7]. Genie 3 (blog post of Google DeepMind dated 5 August 2025) describes generating from text a world that can be navigated in real time at 24 frames per second and 720p, with consistency retained for a few minutes. The same blog lists current limits: the range of actions the agent itself can perform is limited, accurately modeling interactions between multiple agents is still a research challenge, real-world locations cannot be simulated with perfect accuracy, and continuous interaction lasts a few minutes rather than hoursSourced [Source 8].

NVIDIA's Cosmos, on the premise that "Physical AI needs to be trained digitally first", is a platform with a video curation pipeline, pretrained world foundation models and video tokenizers; the paper says the models are released as open weights under permissive licensesSourced [Source 9].

(3) Predicting representations: JEPA

JEPA (Joint-Embedding Predictive Architecture) is a design that predicts representations (features) rather than reconstructing pixels. I-JEPA is a non-generative self-supervised method that predicts the representations of several target blocks of an image from one context block of the same image [Source 4]. V-JEPA is trained only with feature prediction on video (2 million videos) and, the paper reports, yields representations that work for both motion and appearance tasks even with a frozen backbone [Source 5]. V-JEPA 2 was pretrained without actions on over 1 million hours of Internet video, then post-trained into an action-conditioned latent world model (V-JEPA 2-AC) using fewer than 62 hours of unlabelled robot video from the Droid dataset. The paper reports deploying it zero-shot on Franka arms in two labs, picking and placing objects by planning with image goals, without extra trainingSourced [Source 6].

3. Robot foundation models — connecting vision, language and action

After the success of language models, the idea spread of "pretraining one model on large, diverse data and using it for many robots and tasks". A vision-language model that outputs robot actions is called a VLA (vision-language-action) model.

View original SVG, values and source annotations

(1) The thread: from RT-1 and RT-2 to generalist policies

RT-1 (Brohan et al., 2022) built on large-scale data collection with real robots, studied generalization as a function of data size, model size and data diversity, and argued that high-capacity models can absorb diverse robot data [Source 11]. RT-2 co-fine-tunes a web-scale vision-language model on robot trajectory data and web vision-language tasks, expressing actions as text tokens in the same format as language. It reports, over 6,000 evaluation trials, better generalization to novel objects and the ability to interpret commands not in the robot training dataSourced [Source 12].

Open X-Embodiment puts data from 22 robots collected by 21 institutions, showing 527 skills (160,266 tasks), into standardized formats, and reports that RT-X trained on it shows "positive transfer" from other platforms' experience [Source 13]. Octo is an open policy trained on 800,000 trajectories of that data and, it says, can be fine-tuned to new sensors and action spaces within a few hours on consumer GPUs [Source 14]. OpenVLA is a 7-billion-parameter open VLA trained on 970,000 real-robot demonstrations; the paper reports that it beats RT-2-X (55 billion) by 16.5% in absolute task success rate across 29 tasks with 7 times fewer parameters (a value under the paper's conditions)Sourced [Source 15].

(2) How actions are generated: diffusion, flow, text

Ways of producing actions fall into three groups (Fig. 3): as text tokens, as in RT-2; by diffusion, as in Diffusion Policy (Chi et al.); and by flow matching, as in π0. Diffusion Policy reports an average improvement of 46.9% over existing state-of-the-art methods across 12 tasks from 4 robot-manipulation benchmarks [Source 21]. π0 (Physical Intelligence) places a flow-matching architecture on a pretrained VLM, trains on data from several dexterous platforms (single-arm, dual-arm and mobile manipulators), and evaluates on tasks including laundry folding, table cleaning and box assemblySourced [Source 16].

(3) Recent models (developers' technical reports)

All Sourced (descriptions in each technical report). This is not an evaluation comparison table. Robots, tasks, environments and trial counts differ, so these results cannot be lined up to rank models.

4. The gap between simulation and real machines

Training a robot entirely on a real machine takes time and money. So robots are trained in simulation and transferred to the real machine (sim-to-real), but the difference between simulation and reality (the reality gap) is an obstacle. The sources of the gap are sorted below from the papers.

• Randomizing rendering (Tobin et al., 2017): randomizing the simulator's rendering so that "the real world may appear to the model as just another variation". The paper reports that from simulator data with only non-realistic random textures, an object detector accurate to 1.5 cm in the real world could be trained and used for grasping in a cluttered environmentSourced [Source 22]
• Randomizing physical quantities (OpenAI, 2018): in-hand object reorientation on a Shadow Dexterous Hand was trained entirely in a simulation that randomizes many physical properties such as friction coefficients and object appearance, and transferred to the real hand without human demonstrations [Source 23]. A later paper reports solving a Rubik's cube with a humanoid robot hand using automatic domain randomization (ADR)Sourced [Source 24]
• Legged robots: Hwangbo et al. report transferring a neural-network policy trained in simulation to ANYmal, achieving velocity-command tracking, faster running and recovery from falls [Source 25]. Lee et al. then describe a learning-based locomotion controller that enables ANYmal to traverse challenging natural environments [Source 26]Sourced
• A GPU-parallel simulation framework (Isaac Lab): the paper describes it as one framework combining high-fidelity GPU-parallel physics, photorealistic rendering, actuator models, sensor simulation, data collection and domain-randomization toolsSourced [Source 27]

The gap in evaluation: SIMPLER (Li et al.) starts from the point that real-world evaluation of policies does not scale and has reproducibility problems, and builds simulated evaluation environments that mitigate the control and visual disparities. In paired real and simulated evaluations it reports a strong correlation between policy performance in the two, and that behavior modes such as sensitivity to distribution shift are also reflected. This is a result for particular robot setups and does not necessarily hold for every robot and taskSourced [Source 34].

5. Data collection — teleoperation and its neighbours

The "fuel" of robot foundation models is records of robot actions. Unlike the web documents of language models, robot data does not pile up by itself; it has to be gathered by people operating. The main methods are sorted below from the papers.

All Sourced (descriptions in each paper). The way of grouping the methods is this article's own (commentary).

6. How to read success rates (with our calculation)

Robot papers show performance as the success rate of a task. In words familiar to materials and process engineers, it is yield. As with yield, a success rate cannot be interpreted without how many times it was tried and under what conditions.

The scale of evaluation in real papers varies. RT-2 reports 6,000 evaluation trials [Source 12], and RoboArena ranks seven policies with more than 600 real-robot head-to-head episodes (double-blind pairwise comparisons) [Source 35]. RoboArena starts from the point that conventional evaluation, which standardizes fixed tasks, environments and locations, does not readily scale to generalist policies across many tasks and environments, and proposes a crowd-sourced method in which evaluators freely choose the tasks and environments and compare two policies double-blind. Evaluators at seven academic institutions took part using the DROID platform, and the paper reports that it ranks policies more accurately than conventional centralized evaluation and is more scalable, resilient and trustworthySourced [Source 35].

Simulated benchmarks are also used. LIBERO is a lifelong-learning benchmark for robot manipulation that provides four task suites (130 tasks in total) and human-teleoperated demonstrations [Source 32]. CALVIN is a benchmark for learning language-conditioned policies for long-horizon robot manipulation [Source 33]Sourced. Success in simulation, however, does not guarantee success on the real machine (Section 4).

7. Limits, and what could not be confirmed

(1) Performance of current models

The models here (π0, GR00T N1, Gemini Robotics and others) and their success rates are values for the time and settings of each report. We have not covered every latest model as of 15 October 2026 or later evaluations. We have not checked independent third-party reproduction of real-robot results.

(2) Verifying world models

A common standard metric for the "physical accuracy" of world models could not be confirmed. For Genie 3 we did not check details of availability, such as whether it is a limited research preview. The text of Yann LeCun's 2022 position paper (on OpenReview) was not checked because automated retrieval was refused. The explanation of JEPA is limited to what the I-JEPA, V-JEPA and V-JEPA 2 papers describe.

(3) Standards

ISO's standard pages refuse automated retrieval, so numbers, titles, editions, dates and scope were checked from the official-page entries shown in search results; the texts of the standards were not checked. A standard that directly defines how to evaluate the safety of learned policies (robot foundation models) could not be found.

(4) What is not covered

Robot mechanisms, actuators and sensors, and drones are covered in the explainer on drones and robotics; materials discovery and self-driving laboratories in the explainer on materials informatics; autonomous driving in the explainer on autonomous driving and AI. Humanoid products and business plans, robot prices and market size are not covered.

8. How to read evaluations and standards

Evaluating embodied AI has four sides: (1) task success rates (on real robots and in simulation), (2) the quality of a world model's predictions, (3) the quality and quantity of data, and (4) safety. Benchmarks differ by study in tasks, environments and robots, and safety rests on existing standards such as those for industrial robots. An international standard that directly evaluates a learned policy itself could not be found in what we checked.

The organization of the table is this article's own (commentary). The texts of ISO standards are paid and were not checked. Standard stages follow the entries on ISO's official pages shown in search results (checked 15 October 2026).

Three cautions for materials engineers

1. Success rates change with settings and evaluator. Results change with the robot model, task, object placement and evaluator. A developer's own report is not an independent evaluation. A simulated score is no substitute for a real-robot score, and real-robot trials should be read with confidence intervals based on the number of trials (Section 6)
2. Saturation and contamination of benchmarks: in robotics, using the same tasks and the same environment for training and evaluation measures familiarity rather than generalization (a general point that includes an inference). Check that the evaluation environment is not part of training and that the objects and scenes are new. Saturation and contamination of benchmarks in general are covered in this series' explainer on benchmarks and how to read AI evaluations
3. Standards: tell "requirement", "information" and "being drafted" apart. ISO 10218 sets safety requirements for industrial robots, ISO/IEC TR 5469 is a technical report that provides information, and ISO 25785-1 and ISO/IEC TS 22440 are being drafted. "Conforms to the standard" does not mean that a learned policy is safe. Legally binding regulation is covered in this series' explainer on the EU AI Act and others

9. Glossary

10. References (primary sources)

1. Ha D., Schmidhuber J. "World Models", arXiv:1803.10122 (2018) — arxiv.org/abs/1803.10122
2. Hafner D. et al. "Dream to Control: Learning Behaviors by Latent Imagination", arXiv:1912.01603 (2019) — arxiv.org/abs/1912.01603
3. Hafner D. et al. "Mastering diverse control tasks through world models", Nature 640, 647-653 (2025); preprint arXiv:2301.04104, DOI 10.1038/s41586-025-08744-2 — doi.org/10.1038/s41586-025-08744-2
4. Assran M. et al. "Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture", arXiv:2301.08243 (2023) — arxiv.org/abs/2301.08243
5. Bardes A. et al. "Revisiting Feature Prediction for Learning Visual Representations from Video", arXiv:2404.08471 (2024) — arxiv.org/abs/2404.08471
6. Assran M. et al. "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning", arXiv:2506.09985 (2025) — arxiv.org/abs/2506.09985
7. Bruce J. et al. "Genie: Generative Interactive Environments", arXiv:2402.15391 (2024) — arxiv.org/abs/2402.15391
8. Google DeepMind "Genie 3: A new frontier for world models" (blog post dated 5 August 2025). Checked 15 October 2026 — deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models
9. NVIDIA et al. "Cosmos World Foundation Model Platform for Physical AI", arXiv:2501.03575 (2025) — arxiv.org/abs/2501.03575
10. OpenAI "Video generation models as world simulators", technical report (15 February 2024). The official page blocks automated fetching, so the full text was checked in an Internet Archive copy — openai.com/index/video-generation-models-as-world-simulators
11. Brohan A. et al. "RT-1: Robotics Transformer for Real-World Control at Scale", arXiv:2212.06817 (2022) — arxiv.org/abs/2212.06817
12. Brohan A. et al. "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control", arXiv:2307.15818 (2023) — arxiv.org/abs/2307.15818
13. Open X-Embodiment Collaboration "Open X-Embodiment: Robotic Learning Datasets and RT-X Models", arXiv:2310.08864 (2023) — arxiv.org/abs/2310.08864
14. Octo Model Team "Octo: An Open-Source Generalist Robot Policy", arXiv:2405.12213 (2024) — arxiv.org/abs/2405.12213
15. Kim M. J. et al. "OpenVLA: An Open-Source Vision-Language-Action Model", arXiv:2406.09246 (2024) — arxiv.org/abs/2406.09246
16. Black K. et al. (Physical Intelligence) "π0: A Vision-Language-Action Flow Model for General Robot Control", arXiv:2410.24164 (2024; v4 of January 2026) — arxiv.org/abs/2410.24164
17. Physical Intelligence et al. "π0.5: a Vision-Language-Action Model with Open-World Generalization", arXiv:2504.16054 (2025) — arxiv.org/abs/2504.16054
18. NVIDIA et al. "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots", arXiv:2503.14734 (2025) — arxiv.org/abs/2503.14734
19. Gemini Robotics Team "Gemini Robotics: Bringing AI into the Physical World", arXiv:2503.20020 (2025) — arxiv.org/abs/2503.20020
20. Gemini Robotics Team "Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer", arXiv:2510.03342 (2025) — arxiv.org/abs/2510.03342
21. Chi C. et al. "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion", arXiv:2303.04137 (2023) — arxiv.org/abs/2303.04137
22. Tobin J. et al. "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World", arXiv:1703.06907 (2017) — arxiv.org/abs/1703.06907
23. OpenAI et al. "Learning Dexterous In-Hand Manipulation", arXiv:1808.00177 (2018) — arxiv.org/abs/1808.00177
24. OpenAI et al. "Solving Rubik's Cube with a Robot Hand", arXiv:1910.07113 (2019) — arxiv.org/abs/1910.07113
25. Hwangbo J. et al. "Learning agile and dynamic motor skills for legged robots", Science Robotics 4, eaau5872 (2019), DOI 10.1126/scirobotics.aau5872 — doi.org/10.1126/scirobotics.aau5872
26. Lee J. et al. "Learning quadrupedal locomotion over challenging terrain", Science Robotics 5, eabc5986 (2020), DOI 10.1126/scirobotics.abc5986 — doi.org/10.1126/scirobotics.abc5986
27. NVIDIA et al. "Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning", arXiv:2511.04831 (2025) — arxiv.org/abs/2511.04831
28. Zhao T. Z. et al. "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware", arXiv:2304.13705 (2023) — arxiv.org/abs/2304.13705
29. Fu Z. et al. "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation", arXiv:2401.02117 (2024) — arxiv.org/abs/2401.02117
30. Chi C. et al. "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots", arXiv:2402.10329 (2024) — arxiv.org/abs/2402.10329
31. Khazatsky A. et al. "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset", arXiv:2403.12945 (2024) — arxiv.org/abs/2403.12945
32. Liu B. et al. "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning", arXiv:2306.03310 (2023) — arxiv.org/abs/2306.03310
33. Mees O. et al. "CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks", arXiv:2112.03227 (2021) — arxiv.org/abs/2112.03227
34. Li X. et al. "Evaluating Real-World Robot Manipulation Policies in Simulation", arXiv:2405.05941 (2024) — arxiv.org/abs/2405.05941
35. Atreya P. et al. "RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies", arXiv:2506.18123 (2025) — arxiv.org/abs/2506.18123
36. ISO ISO 10218-1:2025, "Robotics — Safety requirements — Part 1: Industrial robots" (Edition 3, February 2025) — iso.org/standard/73933.html
37. ISO ISO 10218-2:2025, "Robotics — Safety requirements — Part 2: Industrial robot applications and robot cells" — iso.org/standard/73934.html
38. ISO ISO/TS 15066:2016, "Robots and robotic devices — Collaborative robots" (February 2016; confirmed in 2022; listed as to be revised) — iso.org/standard/62996.html
39. ISO ISO 13482:2014, "Robots and robotic devices — Safety requirements for personal care robots" (Edition 1, February 2014) — iso.org/standard/53820.html
40. ISO ISO 18646-1:2016, "Robotics — Performance criteria and related test methods for service robots — Part 1: Locomotion for wheeled robots" — iso.org/standard/63127.html
41. ISO ISO 18646-2:2024, "Robotics — Performance criteria and related test methods for service robots — Part 2: Navigation" — iso.org/standard/82643.html
42. ISO ISO 18646-3:2021, "Robotics — Performance criteria and related test methods for service robots — Part 3: Manipulation" — iso.org/standard/73591.html
43. ISO ISO/FDIS 18646-5, "Robotics — Performance criteria and related test methods for service robots — Part 5: Locomotion for legged robots" (final draft stage, per the search-result entry) — iso.org/standard/86850.html
44. ISO ISO/CD 25785-1, "Robotics — Safety requirements for dynamically stable industrial mobile robots (legged, wheeled, or other forms of locomotion) — Part 1: Robots" (committee draft; under development) — iso.org/standard/91469.html
45. ISO ISO 8373:2021, "Robotics — Vocabulary" (Edition 3, November 2021; a revision, ISO/DIS 8373, is in progress) — iso.org/standard/75539.html
46. ISO ISO/IEC TR 5469:2024, "Artificial intelligence — Functional safety and AI systems" (technical report, Edition 1, January 2024) — iso.org/standard/81283.html
47. NIST NIST AI 100-1, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)" (January 2023) — doi.org/10.6028/NIST.AI.100-1

11. Claim-to-source audit

Last updated 15 October 2026. Sources are limited to primary material (arXiv papers, papers in Nature and Science Robotics, developers' technical reports and blog posts, ISO standard pages and official NIST documents). The texts of ISO standards are paid and could not be checked. Because the article includes readings and organization of its own, these are marked as Commentary and kept separate from sourced fact. Latest model performance, independent reproduction, a standard metric for world models, a standard that directly defines the safety of learned policies, and the texts of the standards could not be confirmed and are not stated here. Evaluation values such as success rates are for the settings of each report and do not show which model is better. Figs. 2 to 4 and 6 are vector drawings; Figs. 1 and 5 are AI-generated image slots (none of them shows a real measurement result or a physical robot).