CG-World: A Large-Scale World-State Dataset and Protocol for World Models

# CG-World: A Large-Scale World-State Dataset for World Model Research
Source: [https://arxiv.org/html/2607.26452](https://arxiv.org/html/2607.26452)
###### Abstract

World models require learning the joint distribution of states, actions, events, and observations, whereas existing public video datasets, robotic trajectories, and simulation data typically cover only a subset of these elements\. To address this limitation, we introduce CG\-World \(FigureLABEL:fig:overview\), a world\-state data protocol derived from industrial computer graphics production pipelines, together with a corresponding large\-scale dataset\. The dataset explicitly records intermediate world states, including multimodal semantic information, spatial representations, skeletal and controller states, motion curves, camera and lighting parameters, physics caches, contact events, and multi\-pass rendering outputs\. CG\-World v1 contains approximately 850,000 world\-state segments, each ranging from 1 to 5 seconds\. It separates underlying states, observations, events, relations, and branch metadata, and organizes these heterogeneous records into temporally aligned data units, thereby providing an explicit and comprehensive representation of world\-state information\. To further support intervention\-based learning and counterfactual reasoning, CG\-World introduces a unified branch lineage that categorizes world\-state trajectories into factual trajectories, observational interventions, action interventions, mechanism interventions, and strict counterfactual branches\. For each branch, the intervention target, invariant variables, and alternative outcomes are explicitly recorded\. We evaluate CG\-World on three controlled tasks relevant to world\-model research: geometry\-conditioned video generation, action prediction, and closed\-loop transfer of vision\-language\-action policies\. Experimental results demonstrate that CG\-World provides reusable structured supervision for controlled generation, action modeling, and embodied policy transfer, leading to consistent improvements in model performance\. In future work, we aim to further expand the dataset through continued data collection and community collaboration, with the goal of establishing CG\-World as a shared data infrastructure for general\-purpose world models, Physical AI, and embodied intelligence research\.

Keywords:world models; world\-state data; embodied intelligence; Physical AI; industrial data; counterfactual data; synthetic data

## 1Introduction

World model research is shifting from video generation toward world understanding and prediction\. The field has progressed from compressing observation histories into latent representations to maintaining conditionable, intervenable, and interactive representations in latent space, with the broader goal of jointly modeling state, dynamics, control, and physical regularities\(Dinget al\.[2025](https://arxiv.org/html/2607.26452#bib.bib1)\)\. Kinetics\-700, Ego4D, and Panda\-70M have supported video learning\(Carreiraet al\.[2019](https://arxiv.org/html/2607.26452#bib.bib2); Graumanet al\.[2022](https://arxiv.org/html/2607.26452#bib.bib3); Chenet al\.[2024](https://arxiv.org/html/2607.26452#bib.bib4)\), but scale alone does not guarantee physical understanding\(Motamedet al\.[2026](https://arxiv.org/html/2607.26452#bib.bib5)\)\. Public observational data are dominated by video datasets that rarely provide actions, contacts, and related signals aligned with video, limiting their transfer to decision\-making models for embodied intelligence and autonomous driving\. Robotic datasets such as Open X\-Embodiment, BridgeData V2, and DROID\(Open X\-Embodiment Collaboration[2024](https://arxiv.org/html/2607.26452#bib.bib6); Walkeet al\.[2023](https://arxiv.org/html/2607.26452#bib.bib7); Khazatskyet al\.[2024](https://arxiv.org/html/2607.26452#bib.bib8)\)organize domain\-specific observations, actions, and demonstrations, but generally do not expose solver states, contact graphs, material parameters, or comparable intermediate variables\. Simulation data offer readable states, controllable parameters, and repeatable processes, yet remain the output of particular engines and configurations and are therefore subject to the reality gap, solver bias, and limited cross\-domain generalization\(Hafneret al\.[2025](https://arxiv.org/html/2607.26452#bib.bib9); Liet al\.[2020](https://arxiv.org/html/2607.26452#bib.bib10)\)\.

More consequential than debates over world model architectures is the absence of a data foundation that supports these diverse research directions\. Language and multimodal learning benefit from abundant high\-quality corpora\(Raffelet al\.[2020](https://arxiv.org/html/2607.26452#bib.bib11)\)\. By contrast, public datasets for world models, including video\-generation\-oriented models, usually cover only a subset of visual sequences, camera poses, depth, panoramic observations, and 3D information\. A high\-quality public dataset that jointly aligns 3D scene state, panoramic or multiview observations, depth and geometric ground truth, camera trajectories, and final video within the same spatiotemporal sample remains lacking; multicondition datasets that additionally align actions and physical information are even scarcer\.

Industrial computer graphics production yields not only final videos but also explicit intermediate records of physics, time, and space within CG engines, providing detailed descriptions of world state\. Although these records are not equivalent to real\-world physical ground truth, they numerically describe the motion, interactions, and generative processes underlying the observed output and remain aligned with the final video at both entity and temporal levels\. They are therefore well suited to multimodal process supervision for world models, yet have not been systematically curated into world model training data\.

To address this gap, we follow the precedent of ImageNet\(Denget al\.[2009](https://arxiv.org/html/2607.26452#bib.bib12)\), which organizes images as a hierarchical semantic database\. Centered on spatiotemporal alignment, we establish a hierarchical data\-governance protocol for describing world states and construct a large\-scale dataset and counterfactual branch subset for cross\-task world model training and evaluation\.

## 2Related Work

Real\-world observational data\.ScanNet\(Daiet al\.[2017](https://arxiv.org/html/2607.26452#bib.bib13)\)and Matterport3D\(Changet al\.[2017](https://arxiv.org/html/2607.26452#bib.bib14)\)provide geometry and camera poses from real RGB\-D and 3D scans, offering important support for controllable video generation\. However, these data remain centered on terminal observations and generally lack intermediate physical states strictly aligned with temporal observations\. Genie, GAIA\-1/2, and Cosmos have demonstrated the foundational capabilities of video world models\(Bruceet al\.[2024](https://arxiv.org/html/2607.26452#bib.bib15); Huet al\.[2023](https://arxiv.org/html/2607.26452#bib.bib16); Russellet al\.[2025](https://arxiv.org/html/2607.26452#bib.bib17); NVIDIAet al\.[2025](https://arxiv.org/html/2607.26452#bib.bib18)\)\. Subsequent work such as PhyWorld improves the physical plausibility of video continuations through continued post\-training and physics\-preference optimization, further indicating that terminal video observations alone do not provide directly learnable mechanism supervision\(Zhaoet al\.[2026](https://arxiv.org/html/2607.26452#bib.bib19)\)\.

Sensor\-collected datasets and embodied intelligence\.Open X\-Embodiment aggregates more than one million real robot trajectories in a standardized format\. Models such as RT\-2\(Zitkovichet al\.[2023](https://arxiv.org/html/2607.26452#bib.bib20)\)and OpenVLA\(Kimet al\.[2025b](https://arxiv.org/html/2607.26452#bib.bib21)\)show that jointly training on Internet\-scale vision\-language knowledge and robot behavior data can improve semantic generalization and cross\-task control\. Although these datasets capture real sensors, embodiments, and operational noise, they generally cannot record the complete latent state of the world\. CG\-World is complementary: it supplies dense intermediate\-state supervision that is difficult to obtain through real\-world data collection\.

Synthetic worlds and simulation platforms\.Synthetic data generators and simulators such as Kubric, iGibson, and ThreeDWorld\(Greffet al\.[2022](https://arxiv.org/html/2607.26452#bib.bib22); Liet al\.[2022](https://arxiv.org/html/2607.26452#bib.bib23); Ganet al\.[2021](https://arxiv.org/html/2607.26452#bib.bib24)\)provide controllable environments, rich annotations, physical simulation, and, in some cases, intervention variables\. Yet state protocols and criteria for physical equivalence remain inconsistent across simulators, and different engines or parameter settings may approximate the same physical process with different dynamics\. CG\-World instead derives from delivered production records and includes substantial expert\-authored data unavailable from engine computation alone, such as fluid motion created with 3D particle systems\.

Physics and causality\.IntPhys and PHYRE evaluate intuitive physics and interactive reasoning\(Riochetet al\.[2022](https://arxiv.org/html/2607.26452#bib.bib25); Bakhtinet al\.[2019](https://arxiv.org/html/2607.26452#bib.bib26)\); CLEVRER\(Yiet al\.[2020](https://arxiv.org/html/2607.26452#bib.bib27)\)unifies description, explanation, prediction, and counterfactual reasoning in a video framework; and CausalWorld\(Ahmedet al\.[2021](https://arxiv.org/html/2607.26452#bib.bib28)\)enables explicit interventions on causal variables associated with objects and robots in manipulation environments\. Most visual\-reasoning benchmarks, however, remain limited to discriminative question answering, plausibility judgments, or simplified physical systems and consequently cover a narrow set of application domains\.

In summary, world model training still lacks a standardized data system that systematically aligns semantics, actions, physics, observations, and counterfactuals\. CG\-World addresses this gap by placing explicit intermediate physical states from industrial production, multi\-pass visual observations, and same\-root branch lineages within a shared spatiotemporal data system\.

## 3World\-State Data Protocol

CG\-World defines a logical protocol that separates world\-state elements within aligned samples and maps them to industrial or sensor formats\. This yields a unified, constructible, transformable, and verifiable representation\. Table[2](https://arxiv.org/html/2607.26452#S3.T2)summarizes its seven categories and representative contents\.

Table 2:Core World\-State Data Categories and Representative Contents### 3\.1Logical Schema and Sample Definition

To enable reuse across toolchains, tasks, and models, we decompose the data\-production process and align it with real\-world elements\.

Under this formulation, world\-state data comprise a layered temporal record of a single world evolution process\. Let the canonical time domain of a world segment beT=\{t0,t1,…,tn\}T=\\\{t\_\{0\},t\_\{1\},\\ldots,t\_\{n\}\\\}\. At any timett, the explicit world state is

Wt=\(E,H,P,Xt,Rt,Qt,Πt\)\.W\_\{t\}=\(E,H,P,X\_\{t\},R\_\{t\},Q\_\{t\},\\Pi\_\{t\}\)\.\(1\)
Here,EEis the entity set;HHthe structure set;PPthe static properties;XtX\_\{t\}the dynamic state;RtR\_\{t\}the time\-varying relations;QtQ\_\{t\}the events; andΠt\\Pi\_\{t\}the rules and configurations\. Given an observation functiongmg\_\{m\}for modalitymmand an observation configurationCtmC\_\{t\}^\{m\}, the observation at timettis

Otm=gm​\(Wt;Ctm,ϵtm\)\.O\_\{t\}^\{m\}=g\_\{m\}\\\!\\left\(W\_\{t\};C\_\{t\}^\{m\},\\epsilon\_\{t\}^\{m\}\\right\)\.\(2\)
The modality configurationCtmC\_\{t\}^\{m\}represents sampling mechanisms such as cameras, scanners, or text interfaces, whileϵtm\\epsilon\_\{t\}^\{m\}represents observation noise or encoding perturbations\. Information from each modality is therefore treated as an observation derived from the same world state rather than as the world itself\. With a time\-varying actionata\_\{t\}and an external interventionutu\_\{t\}, world\-state evolution is

Wt\+1=F​\(Wt,at,ut;Θt\)\+ξt\.W\_\{t\+1\}=F\\\!\\left\(W\_\{t\},a\_\{t\},u\_\{t\};\\Theta\_\{t\}\\right\)\+\\xi\_\{t\}\.\(3\)
Here,FFis the state\-transition process jointly induced by the rendering engine, physical simulation, camera system, and related components;Θt\\Theta\_\{t\}denotes the solver, rules, and runtime configuration; andξt\\xi\_\{t\}captures numerical error, stochastic perturbations, or unmodeled factors\. Whenut=∅u\_\{t\}=\\varnothing, the segment records the original or main trajectory; whenut≠∅u\_\{t\}\\neq\\varnothing, it records an intervention branch derived from the same or a shared initial state\.

A complete CG\-World data unit is organized as

Di=\(Wt0:tn,Ot0:tn1:M,At0:tn,Ut0:tn,Bi,ℳi\)\.D\_\{i\}=\\left\(W\_\{t\_\{0\}:t\_\{n\}\},O\_\{t\_\{0\}:t\_\{n\}\}^\{1:M\},A\_\{t\_\{0\}:t\_\{n\}\},U\_\{t\_\{0\}:t\_\{n\}\},B\_\{i\},\\mathcal\{M\}\_\{i\}\\right\)\.\(4\)
In this tuple,Wt0:tnW\_\{t\_\{0\}:t\_\{n\}\}is the world\-state sequence,Ot0:tn1:MO\_\{t\_\{0\}:t\_\{n\}\}^\{1:M\}the multimodal observation sequence,At0:tnA\_\{t\_\{0\}:t\_\{n\}\}the action or control signals,Ut0:tnU\_\{t\_\{0\}:t\_\{n\}\}the intervention variables,BiB\_\{i\}the branch lineage, andℳi\\mathcal\{M\}\_\{i\}the metadata specification, including fixed information such as coordinate systems, units, and rendering configurations\. World model training can therefore extend beyond fitting observations or operational outcomes to jointly learning the next world state:

p​\(Wt\+1,Ot\+1∣W≤t,O≤t,A≤t,U≤t\)\.p\\\!\\left\(W\_\{t\+1\},O\_\{t\+1\}\\mid W\_\{\\leq t\},O\_\{\\leq t\},A\_\{\\leq t\},U\_\{\\leq t\}\\right\)\.\(5\)
![Refer to caption](https://arxiv.org/html/2607.26452v1/figures/figure2.png)Figure 1:Time\-aligned world\-state protocol\.To make the protocol executable, CG\-World declares the semantic and data types, tensor shape, temporal index, units, coordinate system, availability, and provenance of every core field\. Payloads are referenced through standard files or protocols such as OpenUSD\(Pixar Animation Studios[2025](https://arxiv.org/html/2607.26452#bib.bib29)\), RLDS\(Ramoset al\.[2021](https://arxiv.org/html/2607.26452#bib.bib30)\), and MaterialX; the complete schema is provided in the supplementary material\. Each segment is indexed by segment\_id, root\_family\_id, and entity ID and must include a temporal range, entity hierarchy, sensor categories and parameters, and at least one visual observation\. Geometry, animation, physics caches, contacts, depth, normals, and optical flow are optional according to source recoverability\. As illustrated in Figure[1](https://arxiv.org/html/2607.26452#S3.F1), the resulting logical sample layer links world state, actions, state transitions, multimodal observations, relations and events, and causal branches under a unified spatiotemporal index while supporting traceable implementations across industrial tools, simulation platforms, and real\-world acquisition systems\.

### 3\.2Organization Principles and Capability Levels

We adopt three principles for organizing world\-state data:

1. 1\.Separate states, observations, and annotations, preventing a model from treating compressed observations or human labels as the world itself\.
2. 2\.Separate templates, instances, and states, allowing changes to the same instance to be tracked over time\.
3. 3\.Separate continuous states from discrete events while aligning all data to a unified timecode for cross\-modal synchronization\.

We define three capability levels for world model datasets\. L1, the observation\-alignment level, minimally requires entity indices, spatial structure or reconstruction references, camera or sensor parameters, basic observations, unified timecodes, coordinate systems, and units\. L2, the dynamic\-state level, adds action trajectories, entity dynamics, relations, and events\. L3, the causal\-intervention level, further adds intervention variables, invariant fields, branch lineages, outcome variables, and comparable factual\-intervention or factual\-counterfactual pairs\.

### 3\.3Counterfactual and Branch\-Management Protocol

A branch is a world\-state trajectory generated from a traceable world state and its runtime context under specific observation conditions, action inputs, or dynamical mechanisms\. Following the distinction among factual, interventional, and counterfactual reasoning\(Pearl[2009](https://arxiv.org/html/2607.26452#bib.bib31)\), CG\-World divides branches into factual trajectory, observational intervention, action intervention, mechanism intervention, and strict counterfactual branches\.

Factual trajectory branch\.Records the world\-state evolution that occurred and was fully preserved under a given runtime context, action input, and mechanism configuration\.

Observational intervention branch\.Modifies the observation function from the same initial state while keeping the underlying state trajectory unchanged\.

Action intervention branch\.Changes the action or control input applied to the world under a fixed initial world state and runtime context, then records the resulting future states and observations\.

Mechanism intervention branch\.Modifies one or a small number of mechanism variablesIIunder the same initial state to obtain a new world state and subsequent trajectory\.

Strict counterfactual branch\.Conditions on a factual trajectory that occurred and was fully recorded, then applies an alternative intervention to an action, state, or mechanism variable while holding fixed the corresponding exogenous context and nonintervened mechanisms\. CG\-World labels a branch as strict counterfactual data only when it satisfies all four criteria: factual conditioning, preservation of the exogenous context, execution of an alternative intervention, and recording of individual\-level effects\.

## 4CG\-World v1

### 4\.1Data Sources and Processing

CG\-World v1 is a shot\-level world\-state dataset recovered and standardized from industrial CG production projects authorized for research use and redistribution\. We recover temporal correspondences between world states and multiversion rendered observations, using world\-state segments with temporal continuity, persistent entity identities, and strict state\-observation alignment as the basic sample unit\. The pipeline comprises industrial project parsing, cross\-toolchain extraction, and protocol\-based reorganization, as shown in Figure[2](https://arxiv.org/html/2607.26452#S4.F2)\.

![Refer to caption](https://arxiv.org/html/2607.26452v1/figures/figure3.jpg)Figure 2:Processing pipeline from industrial CG projects to CG\-World data packages\.
### 4\.2Data Content, Statistical Units, and Scale

During construction of the CG\-World v1 main dataset, we did not augment the data through procedural scene randomization or resimulation\. All states and observations in v1 originate from saved production records that passed production quality control\.

Figure[3](https://arxiv.org/html/2607.26452#S4.F3)summarizes the full scale, shot\-filtering process, and multimodal coverage of CG\-World v1\. The dataset contains approximately 850,000 valid shot\-level world\-state segments, each 1\-5 seconds long, with a median duration of 2\.6 seconds\. Of these, 433,500 shots have complete multimodal alignment and 510,000 have physics recalculation\. Ninety\-five percent of valid shots include multi\-take action coverage, with 3\-5 take variants per shot\. To prevent leakage across production projects, shared assets, historical versions, and same\-root branches, CG\-World v1 uses grouped hierarchical partitioning\. The final training, validation, and test sets contain 637,500, 127,500, and 85,000 segments, respectively, corresponding to 75%, 15%, and 10% of the valid data\.

![Refer to caption](https://arxiv.org/html/2607.26452v1/figures/figure4.jpg)Figure 3:Data distribution of CG\-World v1\.
### 4\.3Multibranch and Counterfactual Data

Using the branch system introduced in Section 3\.3, CG\-World v1 constructs an intervention and counterfactual subset from desktop container\-manipulation scenes\. The subset is organized by same\-root branch family, as illustrated in Figure[4](https://arxiv.org/html/2607.26452#S4.F4)\. Each family uses one fully recorded factual trajectory as its reference and associates it with alternative branches derived from the same traceable world state\.

![Refer to caption](https://arxiv.org/html/2607.26452v1/figures/figure5.png)Figure 4:Counterfactual data structure in CG\-World\.Approximately 9% of this subset is derived from historical project versions, with the remainder generated procedurally\. Table[3](https://arxiv.org/html/2607.26452#S4.T3)summarizes its principal organizational units, branch scales, intervention targets, and derivable supervision\.

Table 3:Branch and Counterfactual Subset of CG\-World v1
### 4\.4Annotation and Organization

CG\-World v1 uses production\-derived annotation: annotations are recovered from industrial production files rather than manually redrawn as boxes, masks, or textual descriptions over extracted segments\. Textual semantics, structure, actions, physical processes, and observations are produced within the same production chain under human expert supervision\.

### 4\.5Data Rights, Licensing, and Release Plan

The official CG\-World v1 release includes only industrial CG projects that have passed rights review and permit academic research use\. Materials that do not satisfy redistribution requirements are used only for internal statistics or controlled access and are excluded from the public download\.

The complete CG\-World v1 dataset will be released upon paper acceptance and no later than camera\-ready submission\. L1/L2 and selected L3 data will be distributed under a noncommercial academic research license\. Raw L3 data, comprising the complete project files corresponding to world\-state segments, will require a use\-case review and a signed controlled\-access agreement\.

## 5Experiments

We conduct three experiments: observation video generation, multi\-step action prediction, and vision\-language\-action transfer\. Table[4](https://arxiv.org/html/2607.26452#S5.T4)summarizes the evaluated models and initialization checkpoints\.

Table 4:Experimental Models and Initialization Checkpoints### 5\.1Observation Video Generation

We first evaluate whether supervision derived from world state improves observation generation in a video foundation model\. Four settings are compared under the same initial frame, prompt, test clip, random seed, and inference configuration: I2V without explicit control, Union Control LoRA\(Ben\-Yosefet al\.[2026](https://arxiv.org/html/2607.26452#bib.bib35)\), CG\-World single\-modality post\-training, and mixed Depth/Canny post\-training\.

Table 5:Post\-Training Results of LTX\-2\.3 on Observation Video Generation
We use LPIPS\(Zhanget al\.[2018](https://arxiv.org/html/2607.26452#bib.bib36)\)and FVD\(Unterthineret al\.[2018](https://arxiv.org/html/2607.26452#bib.bib37)\)to measure perceptual and distributional distances between generated and real videos\. As shown in Table[5](https://arxiv.org/html/2607.26452#S5.T5), CG\-World geometry\-control supervision yields the best results in both control branches\. Relative to Official Union, LPIPS and FVD decrease by 37\.57% and 38\.75% in the Depth branch and by 33\.05% and 43\.55% in the Canny branch\. These results indicate that strict alignment among same\-source RGB, geometric controls, and temporal information improves perceptual fidelity and distributional consistency in controlled video generation\.

### 5\.2Multi\-Step Action Prediction for Physical AI

The second experiment evaluates whether CG\-World trajectory supervision improves world model action prediction\. The data are organized according to LIBERO\(Liuet al\.[2023](https://arxiv.org/html/2607.26452#bib.bib38)\), and poses use a continuous six\-dimensional rotation representation\(Zhouet al\.[2019](https://arxiv.org/html/2607.26452#bib.bib39)\)\. Using Cosmos3\-Nano as the baseline, we construct three controls with a shared action connector and prediction head\. Action\-Head Only freezes the model backbone and optimizes only the action head and connector; LIBERO\-LoRA post\-trains the backbone on LIBERO; and CG\-World\-LoRA post\-trains it on CG\-World\. Given the current RGB observation and a natural\-language instruction, the model predicts a 16\-step future action chunk\. All three configurations share the same action head, data\-partitioning principles, training budget, and frozen evaluation protocol\.

Table 6:Cosmos3\-Nano Action\-Prediction Results
As shown in Table[6](https://arxiv.org/html/2607.26452#S5.T6), CG\-World\-LoRA achieves the best result on all four action metrics\. Relative to LIBERO\-LoRA, the three continuous\-action errors decrease by 33\.93%, 20\.34%, and 52\.90%, while Gripper F1 improves by 0\.06372 in absolute terms\. These results support the claim that CG\-World provides continuous motion structure beyond independent demonstration trajectories\.

### 5\.3VLA Transfer in Simulation

The third experiment evaluates whether CG\-World provides transferable motion priors for a VLA model and whether those priors improve closed\-loop manipulation under limited target\-domain supervision\. We use OpenVLA\-7B with the OFT continuous\-action regression paradigm\(Kimet al\.[2025a](https://arxiv.org/html/2607.26452#bib.bib40)\)\. Given an RGB observation and task instruction, the model outputs a 7D relative end\-effector action\. We compare post\-training with ManiSkill3\(Taoet al\.[2024](https://arxiv.org/html/2607.26452#bib.bib41)\)and CG\-World\. The CG\-World pre\-adapt \+ OFT condition adds CG\-World action pre\-adaptation before the same target\-domain OFT protocol\. We evaluate in\-domain task success on PickCube, cross\-task generalization on the unseen StackCube task, and decompose failures into Grasp and Lift stages\.

As shown in Table[7](https://arxiv.org/html/2607.26452#S5.T7), CG\-World pretraining raises the in\-domain PickCube success rate from 4\.0% for C2 to 39\.0% and improves StackCube OOD success from 0% for every baseline to 25\.0%\. Its primary benefit is not further reinforcement of already saturated target approach and grasping, but improved post\-grasp lifting, trajectory continuity, object transport, and subsequent action planning\.

Table 7:OpenVLA Simulation Results on Desktop Container\-Manipulation Transfer

## 6Future Work

Future work will use strictly registered real–virtual scene pairs to measure biases in observations, dynamics, and intervention effects; develop error and uncertainty models; and calibrate industrial CG mechanisms and sampling ranges with limited real\-world data\. We will also expand long\-horizon causal chains and dynamical coverage\. Community plugins will map projects to the CG\-World schema, while contributions will require schema, temporal\-consistency, authorization, and provenance checks and receive version and citation metadata under the L1/L2/L3 hierarchy\.

## 7Discussion and Conclusion

### 7\.1Discussion

CG\-World v1 inherits production biases: narrative and shot selection, stylized or corrected motion, engine\- and solver\-specific behavior, incomplete legacy records, and weakened state–observation correspondence from compositing, transparency, or procedural materials\. Its production\-derived states should therefore be used as explicit supervision with documented provenance and operating conditions, not as uncalibrated real\-world physical ground truth\.

### 7\.2Conclusion

CG\-World aligns industrial\-CG states, actions, multimodal observations, and causal branches through a world\-state protocol and large\-scale dataset\. This supervision supports modeling state evolution, observation generation, and intervention outcomes beyond terminal\-pixel fitting, improving controlled video generation, action prediction, and embodied manipulation transfer\.

## References

- O\. Ahmed, F\. Träuble, A\. Goyal, A\. Neitz, Y\. Bengio, B\. Schölkopf, S\. Bauer, and M\. Wüthrich \(2021\)CausalWorld: a robotic manipulation benchmark for causal structure and transfer learning\.InInternational Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p4.1)\.
- A\. Bakhtin, L\. van der Maaten, J\. Johnson, L\. Gustafson, and R\. Girshick \(2019\)PHYRE: a new benchmark for physical reasoning\.InAdvances in Neural Information Processing Systems,Vol\.32\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p4.1)\.
- M\. Ben\-Yosef, T\. Halperin, N\. K\. Korem, M\. Salama, H\. Cain, A\. Joseph, A\. Chen, U\. Jelercic, and O\. Bibi \(2026\)AVControl: efficient framework for training audio\-visual controls\.External Links:2603\.24793Cited by:[§5\.1](https://arxiv.org/html/2607.26452#S5.SS1.p1.1)\.
- J\. Bruce, M\. Dennis, A\. Edwards,et al\.\(2024\)Genie: generative interactive environments\.External Links:2402\.15391Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p1.1)\.
- J\. Carreira, E\. Noland, C\. Hillier, and A\. Zisserman \(2019\)A short note on the Kinetics\-700 human action dataset\.External Links:1907\.06987Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- A\. Chang, A\. Dai, T\. Funkhouser,et al\.\(2017\)Matterport3D: learning from RGB\-D data in indoor environments\.InProceedings of the International Conference on 3D Vision,pp\. 667–676\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p1.1)\.
- T\.\-S\. Chen, A\. Siarohin, W\. Menapace,et al\.\(2024\)Panda\-70M: captioning 70m videos with multiple cross\-modality teachers\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 13320–13331\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- A\. Dai, A\. X\. Chang, M\. Savva, M\. Halber, T\. Funkhouser, and M\. Nießner \(2017\)ScanNet: richly\-annotated 3d reconstructions of indoor scenes\.InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition,pp\. 5828–5839\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p1.1)\.
- J\. Deng, W\. Dong, R\. Socher, L\.\-J\. Li, K\. Li, and L\. Fei\-Fei \(2009\)ImageNet: a large\-scale hierarchical image database\.InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition,pp\. 248–255\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p4.1)\.
- J\. Ding, Y\. Zhang, Y\. Shang,et al\.\(2025\)Understanding world or predicting future? a comprehensive survey of world models\.ACM Computing Surveys58\(3\),pp\. 1–38\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- C\. Gan, J\. Schwartz, S\. Alter,et al\.\(2021\)ThreeDWorld: a platform for interactive multi\-modal physical simulation\.InProceedings of the NeurIPS Track on Datasets and Benchmarks,Vol\.1\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p3.1)\.
- K\. Grauman, A\. Westbury, E\. Byrne,et al\.\(2022\)Ego4D: around the world in 3,000 hours of egocentric video\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 18995–19012\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- K\. Greff, F\. Belletti, L\. Beyer,et al\.\(2022\)Kubric: a scalable dataset generator\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 3749–3761\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p3.1)\.
- Y\. HaCohen, B\. Brazowski, N\. Chiprut,et al\.\(2026\)LTX\-2: efficient joint audio\-visual foundation model\.External Links:2601\.03233Cited by:[Table 4](https://arxiv.org/html/2607.26452#S5.T4.1.2.1.2.1.1)\.
- D\. Hafner, J\. Pasukonis, J\. Ba, and T\. Lillicrap \(2025\)Mastering diverse control tasks through world models\.Nature640,pp\. 647–653\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- A\. Hu, L\. Russell, H\. Yeo, Z\. Murez, G\. Fedoseev, A\. Kendall, J\. Shotton, and G\. Corrado \(2023\)GAIA\-1: a generative world model for autonomous driving\.External Links:2309\.17080Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p1.1)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations,Cited by:[Table 4](https://arxiv.org/html/2607.26452#S5.T4.1.2.1.3.1.1)\.
- A\. Khazatsky, K\. Pertsch, S\. Nair,et al\.\(2024\)DROID: a large\-scale in\-the\-wild robot manipulation dataset\.InProceedings of Robotics: Science and Systems XX,Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- M\. J\. Kim, C\. Finn, and P\. Liang \(2025a\)Fine\-tuning vision\-language\-action models: optimizing speed and success\.External Links:2502\.19645Cited by:[§5\.3](https://arxiv.org/html/2607.26452#S5.SS3.p1.1)\.
- M\. J\. Kim, K\. Pertsch, S\. Karamcheti,et al\.\(2025b\)OpenVLA: an open\-source vision\-language\-action model\.InProceedings of the 8th Conference on Robot Learning,PMLR, Vol\.270,pp\. 2679–2713\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p2.1)\.
- C\. Li, F\. Xia, R\. Martín\-Martín,et al\.\(2022\)iGibson 2\.0: object\-centric simulation for robot learning of everyday household tasks\.InProceedings of the 5th Conference on Robot Learning,PMLR, Vol\.164,pp\. 455–465\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p3.1)\.
- M\. Li, M\. Yang, F\. Liu, X\. Chen, Z\. Chen, and J\. Wang \(2020\)Causal world models by unsupervised deconfounding of physical dynamics\.External Links:2012\.14228Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- B\. Liu, Y\. Zhu, C\. Gao, Y\. Feng, Q\. Liu, Y\. Zhu, and P\. Stone \(2023\)LIBERO: benchmarking knowledge transfer for lifelong robot learning\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[§5\.2](https://arxiv.org/html/2607.26452#S5.SS2.p1.1)\.
- S\. Motamed, L\. Culp, K\. Swersky, P\. Jaini, and R\. Geirhos \(2026\)Do generative video models understand physical principles?\.InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision,pp\. 948–958\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- NVIDIA, N\. Agarwal, A\. Ali,et al\.\(2025\)Cosmos world foundation model platform for physical AI\.External Links:2501\.03575Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p1.1)\.
- NVIDIA \(2026\)Cosmos 3: omnimodal world models for physical AI\.External Links:2606\.02800Cited by:[Table 4](https://arxiv.org/html/2607.26452#S5.T4.1.3.2.2.1.1)\.
- Open X\-Embodiment Collaboration \(2024\)Open x\-embodiment: robotic learning datasets and RT\-X models\.InProceedings of the IEEE International Conference on Robotics and Automation,pp\. 6892–6903\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- J\. Pearl \(2009\)Causality: models, reasoning, and inference\.2 edition,Cambridge University Press\.Cited by:[§3\.3](https://arxiv.org/html/2607.26452#S3.SS3.p1.1)\.
- Pixar Animation Studios \(2025\)Universal scene description \(OpenUSD\)\.Note:OpenUSD Official Documentation,https://openusd\.orgCited by:[§3\.1](https://arxiv.org/html/2607.26452#S3.SS1.p13.1)\.
- C\. Raffel, N\. Shazeer, A\. Roberts,et al\.\(2020\)Exploring the limits of transfer learning with a unified text\-to\-text transformer\.Journal of Machine Learning Research21\(140\),pp\. 1–67\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p2.1)\.
- S\. Ramos, S\. Girgin, L\. Hussenot,et al\.\(2021\)RLDS: an ecosystem to generate, share and use datasets in reinforcement learning\.External Links:2111\.02767Cited by:[§3\.1](https://arxiv.org/html/2607.26452#S3.SS1.p13.1)\.
- R\. Riochet, M\. Y\. Castro, M\. Bernard, A\. Lerer, R\. Fergus, V\. Izard, and E\. Dupoux \(2022\)IntPhys 2019: a benchmark for visual intuitive physics understanding\.IEEE Transactions on Pattern Analysis and Machine Intelligence44\(9\),pp\. 5016–5025\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p4.1)\.
- L\. Russell, A\. Hu, L\. Bertoni, G\. Fedoseev, J\. Shotton, E\. Arani, and G\. Corrado \(2025\)GAIA\-2: a controllable multi\-view generative world model for autonomous driving\.External Links:2503\.20523Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p1.1)\.
- S\. Tao, F\. Xiang, A\. Shukla,et al\.\(2024\)ManiSkill3: GPU parallelized robotics simulation and rendering for generalizable embodied AI\.External Links:2410\.00425Cited by:[§5\.3](https://arxiv.org/html/2607.26452#S5.SS3.p1.1)\.
- T\. Unterthiner, S\. van Steenkiste, K\. Kurach, R\. Marinier, M\. Michalski, and S\. Gelly \(2018\)Towards accurate generative models of video: a new metric & challenges\.External Links:1812\.01717Cited by:[§5\.1](https://arxiv.org/html/2607.26452#S5.SS1.p2.1)\.
- H\. R\. Walke, K\. Black, T\. Z\. Zhao,et al\.\(2023\)BridgeData V2: a dataset for robot learning at scale\.InProceedings of the 7th Conference on Robot Learning,PMLR, Vol\.229,pp\. 1723–1736\.Cited by:[§1](https://arxiv.org/html/2607.26452#S1.p1.1)\.
- K\. Yi, C\. Gan, Y\. Li, P\. Kohli, J\. Wu, A\. Torralba, and J\. B\. Tenenbaum \(2020\)CLEVRER: collision events for video representation and reasoning\.InInternational Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p4.1)\.
- R\. Zhang, P\. Isola, A\. A\. Efros, E\. Shechtman, and O\. Wang \(2018\)The unreasonable effectiveness of deep features as a perceptual metric\.InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition,pp\. 586–595\.Cited by:[§5\.1](https://arxiv.org/html/2607.26452#S5.SS1.p2.1)\.
- P\. Zhao, J\. Lin, T\. Rupprecht,et al\.\(2026\)PhyWorld: physics\-faithful world model for video generation\.External Links:2605\.19242Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p1.1)\.
- Y\. Zhou, C\. Barnes, J\. Lu, J\. Yang, and H\. Li \(2019\)On the continuity of rotation representations in neural networks\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 5745–5753\.Cited by:[§5\.2](https://arxiv.org/html/2607.26452#S5.SS2.p1.1)\.
- B\. Zitkovich, T\. Yu, S\. Xu,et al\.\(2023\)RT\-2: vision\-language\-action models transfer web knowledge to robotic control\.InProceedings of the 7th Conference on Robot Learning,PMLR, Vol\.229,pp\. 2165–2183\.Cited by:[§2](https://arxiv.org/html/2607.26452#S2.p2.1)\.