RoboPhys-3D: A Comprehensive Embodied World Model Evaluation via 3D Reconstruction

Tianyi Wang; Jiazhou Chen; Yiming Xu; Xiangyu Li; Tianyi Zeng; Chih-Hsien Chou; Ning Lu; Liang Peng

cs.ROAug 28, 2026

RoboPhys-3D: A Comprehensive Embodied World Model Evaluation via 3D Reconstruction

Authors: Tianyi Wang, Jiazhou Chen, Yiming Xu, Xiangyu Li, Tianyi Zeng, Chih-Hsien Chou, Ning Lu, Liang Peng,

Organizations: The University of Texas at Austin · Purdue University · Futurewei Technologies, Inc.

Abstract

Video world models increasingly serve as data engines, action planners, and simulators for embodied AI, but conventional embodied world model (EWM) benchmarks lack a unified 3D-grounded protocol for establishing whether generated rollouts preserve the underlying 3D scene state or translate into executable actions. We introduce RoboPhys-3D, a 3D-grounded EWM benchmark built on RoboTwin 2.0, covering 50 manipulation tasks across four regimes, with 5,000 episodes and 25,000 multi-view ground-truth videos. A defining feature of RoboPhys-3D is that generated and ground-truth videos are processed through the same 3D reconstruction pipeline, enabling reconstruction-induced error to be distinguished from generation-induced error. The RoboPhys-3D benchmark organizes 50 complementary metrics into 18 sub-dimensions across four levels: pixel-level fidelity, 3D geometry consistency, state-level understanding, and task-level completeness. We further introduce Average Full Score, a hierarchical score averaging all 50 metrics for comprehensive evaluation, and RoboPhyscore, a compact task-aligned score averaging the metrics most strongly correlated with task success. Among the four representative video world models, Cosmos 3 achieves the highest RoboPhyscore (0.6330, 92.7% of ground truth), while state- and execution-grounded metrics reveal substantial failures that perceptual and vision-language model-based judgments fail to capture. RoboPhyscore further exhibits strong agreement with human evaluation (Pearson r = 0.9761 and Spearman \r{ho} = 0.8962), demonstrating the importance of grounded, execution-aware evaluation for EWM capability.

Figures & tables

Metrics	WBench [ 79 ]	ESI-Bench [ 18 ]	WorldArena [ 59 ]	RoboWM-Bench [ 26 ]	EZSBench [ 8 ]	Wow,wo,val [ 10 ]	WorldModelBench [ 31 ]	EWMBench [ 81 ]	World-in-World [ 83 ]	Ours
Venue & Year	arXiv 26’	arXiv 26’	CVPRW 26’	CVPRW 26’	arXiv 26’	arXiv 26’	NeurIPS 25’	arXiv 25’	ICLR 26’	/
Pixel-Level Fidelity
IQ-1	✓	×	✓	×	✓	✓	✓	×	✓	✓
AQ	✓	×	✓	×	✓	×	×	×	✓	✓
DS	×	×	✓	×	×	✓	×	×	×	✓
MQ	✓	×	✓	×	✓	✓	✓	✓	×	✓

Table 1: Comparison of the RoboPhys-3D benchmark (ours) to other existing embodied world model (EWM) benchmarks. (1) Pixel-level fidelity: IQ-1 = Image Quality, AQ = Aesthetic Quality, DS = Distribution Similarity, MQ = Motion Quality, CC = Content Consistency. (2) 3D geometry consistency: RBG = Reconstruction-Based Geometry, DC = Depth Consistency, CTG = Camera Trajectory Geometry, RC = Roundtrip Consistency. (3) State-level understanding: OL = Object Localization, TA = Trajectory Accuracy, IQ-2 = Interaction Quality, EAA = Event and Action Adherence, PLA = Physical Law Adherence, SA = Semantic Adherence. (4) Task-level completeness: IF = Instruction Following, CS = Collision Safety, TS = Task Success.

Figure 1: Overview of the proposed RoboPhys-3D benchmark.

Figure 2: Overall results of Average Full Score (AFS) and RoboPhyscore. (a) RoboPhyscore for different video world models and reconstruction methods. (b) AFS for different video world models and reconstruction methods.

Metrics	Ground Truth	General World Model		Embodied World Model
		Wan	CogVideoX	Cosmos	RoboDreamer
Pixel-Level Fidelity (/5)	3.1300	2.7249	2.1236	2.9329	2.1359
IQ-1 (/6)	4.3156	3.5108	3.0803	3.6128	3.2018
AQ (/4)	1.3323	1.6047	1.3855	1.5019	1.0521
DS (/1)	0.8142	0.3602	0.1505	0.6525	0.0898
MQ (/4)	2.5848	3.1620	2.0245	2.8713	2.6996

Table 2: Detailed quantitative results of the RoboPhys-3D benchmark for four representative video world models.

Metrics	Ground Truth	Feed-Forward Estimator		Dynamic Representation
		VGGT	VGGT- $\Omega$	4DGS	4C4D
Pixel-Level Fidelity (/5)	2.7784	2.6583	2.6806	2.4376	2.4923
IQ-1 (/6)	3.7084	3.4867	3.5443	3.5562	3.4258
AQ (/4)	1.6312	1.4247	1.4841	1.0956	1.2409
DS (/1)	0.4802	0.4646	0.4664	0.3218	0.3340
MQ (/4)	2.6947	2.6530	2.6409	2.6261	2.7275

Table 3: Detailed quantitative results of the RoboPhys-3D benchmark for four representative reconstruction methods.

Method	Ground Truth	VGGT	VGGT- $\Omega$	4DGS	4C4D
Ground Truth	0.8361	0.7444	0.7592	0.4660	0.6088
Wan	0.5782	0.5559	0.5609	0.4816	0.5029
CogVideoX	0.3473	0.3269	0.3366	0.3387	0.3322
Cosmos	0.6800	0.6649	0.6705	0.5530	0.5965
RoboDreamer	0.3503	0.3519	0.3549	0.3453	0.3434

Table 4: RoboPhyscore for each video world model and each reconstruction method.

Method	Task 1	Task 2	Task 3	Task 4
Ground Truth	0.7117	0.7033	0.6702	0.6464
Wan	0.5596	0.5778	0.5306	0.4756
CogVideoX	0.3424	0.3453	0.3290	0.3285
Cosmos	0.6546	0.6648	0.6245	0.5881
RoboDreamer	0.3452	0.3779	0.3450	0.3287

Table 5: RoboPhyscore for different task categories. Task 1 is single-arm w/o target. Task 2 is single-arm w/ target. Task 3 is dual-arm w/o target. Task 4 is dual-arm w/ target.

Method	Instruction	Previous Prompt	Current Prompt
Wan	0.4962	0.5474	0.5782
Cosmos	0.6519	0.6541	0.6800

Table 6: RoboPhyscore for different scene annotations.

Method	MAE $\downarrow$	Task Success (Action Planner) $\uparrow$	Task Success (Data Engine) $\uparrow$
J-IDM	0.0744	0.7575	0.4875
MIDM	0.0273	0.9000	0.4750
DreamGen	0.0112	0.9825	0.5875

Table 7: Performance of different inverse dynamic models.

Figure 3: Cross-dimension correlation analysis. (a) Pearson correlation among 18 sub-dimensions. (b) Pearson correlation among four levels. (c) Pearson correlation between RoboPhyscore and human evaluation.

Figure 4: Qualitative examples of Adjust Bottle task (different world models + ground truth + view-aware prompt).

Appendix figures & tables56 assets

Supplementary material from the paper’s appendix.

Appendix

Component	RoboPhys-3D
Simulation Platform	RoboTwin 2.0
Manipulation Regimes	4
Tasks	50
Episodes per Task	100
Camera Views per Episode	5
Single-Arm w/o Target Task	11

Appendix

Table 8: Key statistics of the RoboPhys-3D dataset.

	Without Target	With Target
Single-Arm	(11 Tasks) Adjust Bottle, Click Alarmclock, Click Bell, Press Stapler, Turn Switch, Shake Bottle, Shake Bottle Horizontally, Open Laptop, Open Microwave, Rotate QRcode, Move Playingcard Away	(21 Tasks) Move Can Pot, Move Pillbottle Pad, Move Stapler Pad, Place A2B Left, Place A2B Right, Place Container Plate, Place Empty Cup, Place Fan, Place Mouse Pad, Place Object Scale, Place Object Stand, Place Phone Stand, Place Shoe, Stack Blocks Two, Stack Blocks Three, Stack Bowls Two, Stack Bowls Three, Blocks Ranking RGB, Blocks Ranking Size, Beat Block Hammer, Stamp Seal
Dual-Arm	(5 Tasks) Grab Roller, Lift Pot, Pick Dual Bottles, Pick Diverse Bottles, Handover Mic	(13 Tasks) Put Object Cabinet, Handover Block, Hanging Mug, Place Bread Skillet, Place Dual Shoes, Place Burger Fries, Place Cans Plasticbox, Place Can Basket, Place Object Basket, Put Bottles Dustbin, Scan Object, Place Bread Basket, Dump Bin Bigbin

Appendix

Table 9: Task taxonomy of the 50 RoboTwin 2.0 manipulation tasks in RoboPhys-3D.

Method	Paradigm	Representation	Pose Input	Temporal Modeling	Model / Scene Size
Pose-Free Feed-Forward Estimator
VGGT	Feed-Forward	Point/Depth Maps	Not Required	None	1.26 B Parameters
VGGT- $\Omega$	Feed-Forward	Depth Maps + Cameras	Not Required	None	1.14 B Parameters
Dynamic 4D Representation
4DGS	Per-Scene Optimization	3D Gaussians + 4D Deformation Field	Required	Time-Conditioned Deformation	0.18 M Gaussians/Scene
4C4D	Per-Scene Optimization	4D Gaussians	Required	Explicit 4D Dynamics	4.28 M Gaussians/Scene

Appendix

Table 10: Characteristics of the evaluated 3D/4D reconstruction methods.

Method	Architecture	Conditioning	Evaluation Resolution	Frames	Inference (s/video)	Parameters
General World Model
Wan-2.2-I2V-A14B	MoE DiT	Video + Text	1280 $\times$ 720	81	299.0	15.17 B
CogVideoX1.5-5B	3D DiT	Video + Text	1280 $\times$ 768	81	191.0	10.55 B
Embodied World Model
Cosmos-3-Nano	MoT DiT	Video + Text	1280 $\times$ 720	81	33.4	16.40 B
RoboDreamer	Spatiotemporal U-Net	Video + Text	1280 $\times$ 768	81	32.0	340.50 M

Appendix

Table 11: Characteristics of the evaluated video world models.

Method	Vision Encoder	Feature Dimensions	Encoder Parameters	Action Module	Prediction Mechanism	Module Parameters
J-IDM	DINOv2 ViT-L	1024	303.3 M	DPT	Jacobian Field + Pseudoinverse	32.7 M
MIDM	ResNet-50	2048	23.6 M	Mask Predictor + Regressor	Masked Action Regression	0.03 M
DreamGen	SigLIP2-L	1024	316.0 M	DiT	Action Flow Matching	321.4 M

Appendix

Table 12: Detailed model parameters of inverse dynamic models.

Metric	Direction	Upper	Lower	Bound Source
Flow Score	$\uparrow$	8.9414	0.0531	WorldArena
Motion Smoothness	$\uparrow$	2.6413	0.0000	WorldArena
Photometric Consistency	$\uparrow$	6.7899	0.1257	WorldArena
Depth Accuracy	$\downarrow$	4.3711	0.2228	WorldArena
Style Consistency	$\downarrow$	0.0070	0.0000	WorldScore
PSNR	$\uparrow$	50.0000	0.0000	WoW-World-Eval

Appendix

Table 13: Bounds used for score normalization.

Metric	Grounding
PSNR (IQ-1)	Video-Grounded
SSIM (IQ-1)	Video-Grounded
FVD (IQ-1)	Video-Grounded
LPIPS (IQ-1)	Video-Grounded
JEPA Similarity (DS)	Video-Grounded
Depth Accuracy (DC)	Video-Grounded

Appendix

Table 14: Metrics evaluated with an external reference.

Metric
MUSIQ (IQ-1)
Inception Score (IQ-1)
LAION AP (AQ)
HPSv3 (AQ)
CLIP-AP (AQ)
CLIP-IQA (AQ)

Appendix

Table 15: Metrics evaluated without any external reference.

Figure 6: Detailed results of the RoboPhys-3D benchmark for four representative video world models.

Figure 7: Detailed results of the RoboPhys-3D benchmark for four representative reconstruction methods.

Figure 8: Detailed results of the proposed RoboPhyscore. (a) RoboPhyscore for representative video world models. (b) RoboPhyscore for representative reconstruction methods.

Figure 9: Pearson correlation among 50 metrics.

Threshold	Pearson $r$ $\uparrow$	Spearman $\rho$ $\uparrow$	$R^{2}$ $\uparrow$	MAE $\downarrow$	# of Metrics
Human Alignment
0.5	0.9592	0.9200	0.9201	0.0419	19
0.6	0.9649	0.9023	0.9310	0.0404	12
0.7	0.9761	0.8962	0.9528	0.0317	8
0.8	0.9454	0.9692	0.8938	0.0499	1
0.9	N/A	N/A	N/A	N/A	0

Appendix

Table 16: Parameter sensitivity of RoboPhyscore under different threshold settings.

Figure 10: Qualitative examples of Move Pillbottle Pad task (Cosmos + different reconstruction methods + view-aware prompt).

Figure 11: Qualitative examples of Grab Roller task (Wan + ground truth + different prompts).

Figure 12: Qualitative examples of Put Object Cabinet task (ground truth + different inverse dynamic models).

Average Full Score (/4)
Method	Ground Truth	VGGT	VGGT- $\Omega$	4DGS	4C4D
Ground Truth	3.1563	2.9475	2.9777	2.4591	2.6492
Wan	2.7320	2.5793	2.6126	2.4454	2.4878
CogVideoX	2.2774	2.2081	2.2573	2.2666	2.2491
Cosmos	2.9024	2.7977	2.8234	2.6321	2.7431
RoboDreamer	2.1318	2.1134	2.1280	2.1292	2.1206

Appendix

Table 17: Different unified scores for each video world model and each reconstruction method.

Unified Score	Pearson $r$ $\uparrow$	Spearman $\rho$ $\uparrow$	$R^{2}$ $\uparrow$	MAE $\downarrow$
Average Full Score	0.9746	0.9715	0.9499	0.0284
RoboPhyscore	0.9761	0.8962	0.9528	0.0317
WorldModelBench	0.8090	0.8200	0.6545	0.0868
EWMBench	0.6417	0.8669	0.4117	0.1282
WorldArena	0.9263	0.8046	0.8580	0.0497

Appendix

Table 18: Agreement between different unified scores and human evaluation.

Action Planner
Unified Score	Pearson $r$ $\uparrow$	Spearman $\rho$ $\uparrow$	$R^{2}$ $\uparrow$	MAE $\downarrow$
Average Full Score	0.9503	0.9525	0.9031	0.0732
RoboPhyscore	0.9551	0.8586	0.9122	0.0749
WorldModelBench	0.7217	0.7359	0.5209	0.1992
EWMBench	0.6505	0.8175	0.4231	0.2279
WorldArena	0.8552	0.7475	0.7313	0.1500

Appendix

Table 19: Agreement between different unified scores and task success evaluation.

Method	MUSIQ	PSNR	SSIM	LPIPS	FVD	Inception Score	IQ-1
RoboTwin							4.3156/6
GT	0.6135	1.0000	1.0000	1.0000	1.0000	0.2240	4.8375
VGGT	0.5579	0.6912	0.9711	0.9262	0.9973	0.1922	4.3359
VGGT- $\Omega$	0.5765	0.7188	0.9758	0.9501	0.9979	0.1853	4.4044
4DGS	0.3907	0.4892	0.9052	0.8241	0.9446	0.5588	4.1126
4C4D	0.4122	0.5576	0.9560	0.8413	0.9198	0.2007	3.8876

Appendix

Table 20: Image quality score comparison of different world models.

Method	MUSIQ	PSNR	SSIM	LPIPS	FVD	Inception Score	IQ-1
Wan							3.5599/6
GT+Instruction	0.6345	0.2264	0.7645	0.6596	0.7741	0.3907	3.4498
GT+Previous Prompt	0.6523	0.2326	0.7768	0.6688	0.8009	0.4075	3.5389
GT+Current Prompt	0.6596	0.2366	0.7814	0.6782	0.8118	0.5235	3.6911
Cosmos							3.7059/6
GT+Instruction	0.6237	0.3436	0.8526	0.8054	0.9296	0.1412	3.6961

Appendix

Table 21: Image quality score comparison of different prompts.

Method	LAION-AP	HPSv3	CLIP-AP	CLIP-IQA	AQ
RoboTwin					1.3323/4
GT	0.4192	0.2258	0.7231	0.4050	1.7731
VGGT	0.4234	0.1655	0.5940	0.3260	1.5089
VGGT- $\Omega$	0.4250	0.1898	0.6220	0.3480	1.5848
4DGS	0.3351	0.0068	0.1985	0.2630	0.8034
4C4D	0.3607	0.0257	0.3150	0.2897	0.9911

Appendix

Table 22: Aesthetic quality score comparison of different world models.

Method	LAION-AP	HPSv3	CLIP-AP	CLIP-IQA	AQ
Wan					1.8578/4
GT+Instruction	0.4138	0.2336	0.7104	0.4724	1.8302
GT+Previous Prompt	0.4171	0.2479	0.7130	0.4768	1.8548
GT+Current Prompt	0.4183	0.2580	0.7231	0.4889	1.8883
Cosmos					1.6861/4
GT+Instruction	0.3989	0.2147	0.6502	0.4024	1.6662

Appendix

Table 23: Aesthetic quality score comparison of different prompts.

Method	JEPA Similarity (DS)
RoboTwin	0.8142/1
GT	1.0000
VGGT	0.9683
VGGT- $\Omega$	0.9664
4DGS	0.5328
4C4D	0.6036

Appendix

Table 24: Distribution similarity score comparison of different world models.

Method	JEPA Similarity (DS)
Wan	0.3746/1
GT+Instruction	0.3120
GT+Previous Prompt	0.4033
GT+Current Prompt	0.4086
Cosmos	0.7478/1
GT+Instruction	0.7405

Appendix

Table 25: Distribution similarity score comparison of different prompts.

Method	Dynamic Degree	Flow Score	Motion Smoothness	Temporal Flickering	MQ
RoboTwin					2.5848/4
GT	0.5786	0.3042	0.7651	0.9834	2.6313
VGGT	0.5743	0.2966	0.7627	0.9834	2.6170
VGGT- $\Omega$	0.5746	0.2945	0.7637	0.9834	2.6162
4DGS	0.5175	0.2119	0.6436	0.9855	2.3585
4C4D	0.6196	0.4206	0.6796	0.9814	2.7012

Appendix

Table 26: Motion quality score comparison of different world models.

Method	Dynamic Degree	Flow Score	Motion Smoothness	Temporal Flickering	MQ
Wan					3.1164/4
GT+Instruction	0.7093	0.5808	0.8421	0.9675	3.0997
GT+Previous Prompt	0.6986	0.5203	0.8567	0.9686	3.0442
GT+Current Prompt	0.7516	0.6043	0.8792	0.9702	3.2053
Cosmos					2.8610/4
GT+Instruction	0.5930	0.4964	0.7912	0.9725	2.8531

Appendix

Table 27: Motion quality score comparison of different prompts.

Method	Subject	Background	Style	Photometric	Semantics	CC
RoboTwin						3.0863/5
GT	0.8167	0.8725	0.9734	0.2226	0.2113	3.0965
VGGT	0.8155	0.8558	0.9843	0.2392	0.2173	3.1121
VGGT- $\Omega$	0.8152	0.8563	0.9827	0.2397	0.2173	3.1112
4DGS	0.8199	0.8954	0.9872	0.1823	0.1508	3.0356
4C4D	0.7965	0.8726	0.9862	0.2008	0.2200	3.0761

Appendix

Table 28: Content consistency score comparison of different world models.

Method	Subject	Background	Style	Photometric	Semantics	CC
Wan						2.9504/5
GT+Instruction	0.8124	0.8880	0.9743	0.0912	0.1762	2.9421
GT+Previous Prompt	0.7959	0.8816	0.9698	0.0993	0.1843	2.9309
GT+Current Prompt	0.7896	0.8749	0.9628	0.1012	0.2497	2.9782
Cosmos						2.9173/5
GT+Instruction	0.7937	0.8331	0.9724	0.0940	0.1713	2.8645

Appendix

Table 29: Content consistency score comparison of different prompts.

Method	Fidelity	Stability	Reprojection Geometric	Reprojection Photometric	RBG
RoboTwin					1.8689/4
GT	0.5386	0.3960	0.9339	0.1974	2.0659
VGGT	0.2993	0.4873	0.9059	0.2017	1.8942
VGGT- $\Omega$	0.3043	0.4881	0.9081	0.2010	1.9015
4DGS	0.0891	0.4912	0.9550	0.2244	1.7597
4C4D	0.0969	0.4714	0.9676	0.1875	1.7234

Appendix

Table 30: Reconstruction-based geometry score comparison of different world models.

Method	Fidelity	Stability	Reprojection Geometric	Reprojection Photometric	RBG
Wan					2.0835/4
GT+Instruction	0.5378	0.3883	0.9722	0.2288	2.1271
GT+Previous Prompt	0.6071	0.3305	0.9646	0.1811	2.0833
GT+Current Prompt	0.6339	0.2968	0.9576	0.1517	2.0400
Cosmos					1.9700/4
GT+Instruction	0.4980	0.3535	0.9232	0.2034	1.9781

Appendix

Table 31: Reconstruction-based geometry score comparison of different prompts.

Method	Depth Accuracy	Perspectivity	DC
RoboTwin			1.7086/2
GT	1.0000	0.8965	1.8965
VGGT	1.0000	0.8560	1.8560
VGGT- $\Omega$	1.0000	0.8725	1.8725
4DGS	0.7982	0.6180	1.4162
4C4D	0.7256	0.7760	1.5016

Appendix

Table 32: Depth consistency score comparison of different world models.

Method	Depth Accuracy	Perspectivity	DC
Wan			1.5715/2
GT+Instruction	0.6615	0.8005	1.4620
GT+Previous Prompt	0.6909	0.8720	1.5629
GT+Current Prompt	0.7116	0.9780	1.6896
Cosmos			1.7879/2
GT+Instruction	0.9081	0.8285	1.7366

Appendix

Table 33: Depth consistency score comparison of different prompts.

Method	Camera Control (CTG)
RoboTwin	0.8687/1
GT	0.8724
VGGT	0.9050
VGGT- $\Omega$	0.9021
4DGS	0.9476
4C4D	0.7166

Appendix

Table 34: Camera trajectory geometry score comparison of different world models.

Method	Camera Control (CTG)
Wan	0.8820/1
GT+Instruction	0.8759
GT+Previous Prompt	0.8831
GT+Current Prompt	0.8871
Cosmos	0.9053/1
GT+Instruction	0.8946

Appendix

Table 35: Camera trajectory geometry score comparison of different prompts.

Method	Spatial Consistency (RC)
RoboTwin	0.7242/1
GT	0.7160
VGGT	0.7130
VGGT- $\Omega$	0.7121
4DGS	0.7455
4C4D	0.7344

Appendix

Table 36: Roundtrip consistency score comparison of different world models.

Method	Spatial Consistency (RC)
Wan	0.7277/1
GT+Instruction	0.7468
GT+Previous Prompt	0.7256
GT+Current Prompt	0.7108
Cosmos	0.7136/1
GT+Instruction	0.7117

Appendix

Table 37: Roundtrip consistency score comparison of different prompts.

Method	Average Precision	Average Recall	OL
RoboTwin			1.3527/2
GT	1.0000	1.0000	2.0000
VGGT	0.7228	0.7774	1.5002
VGGT- $\Omega$	0.7673	0.8143	1.5816
4DGS	0.2545	0.3826	0.6371
4C4D	0.4713	0.5731	1.0444

Appendix

Table 38: Object localization score comparison of different world models.

Method	Average Precision	Average Recall	OL
Wan			0.4801/2
GT+Instruction	0.1609	0.2944	0.4553
GT+Previous Prompt	0.1788	0.2970	0.4758
GT+Current Prompt	0.1945	0.3147	0.5092
Cosmos			0.8957/2
GT+Instruction	0.3892	0.5045	0.8937

Appendix

Table 39: Object localization score comparison of different prompts.

Method	Mean Trajectory Length	HSD	nDTW	DYN	TA
RoboTwin					2.0088/4
GT	0.7295	1.0000	1.0000	1.0000	3.7295
VGGT	0.7298	0.3166	0.5675	0.1042	1.7181
VGGT- $\Omega$	0.7292	0.3337	0.6221	0.1411	1.8261
4DGS	0.7308	0.2383	0.3767	0.0935	1.4393
4C4D	0.7298	0.1842	0.3999	0.0172	1.3311

Appendix

Table 40: Trajectory accuracy score comparison of different world models.

Method	Mean Trajectory Length	HSD	nDTW	DYN	TA
Wan					1.4444/4
GT+Instruction	0.7319	0.2356	0.3269	0.1347	1.4291
GT+Previous Prompt	0.7282	0.2491	0.3358	0.1308	1.4439
GT+Current Prompt	0.7209	0.2560	0.3403	0.1429	1.4601
Cosmos					1.6547/4
GT+Instruction	0.7485	0.3090	0.4694	0.1174	1.6443

Appendix

Table 41: Trajectory accuracy score comparison of different prompts.

Method	Interaction Score (IQ-2)
RoboTwin	0.6216/1
GT	0.6955
VGGT	0.6555
VGGT- $\Omega$	0.6710
4DGS	0.4890
4C4D	0.5970

Appendix

Table 42: Interaction quality score comparison of different world models.

Method	Interaction Score (IQ-2)
Wan	0.6470/1
GT+Instruction	0.5370
GT+Previous Prompt	0.6695
GT+Current Prompt	0.7345
Cosmos	0.6610/1
GT+Instruction	0.6275

Appendix

Table 43: Interaction quality score comparison of different prompts.

Method	Event Editing	Subject Action	EAA
RoboTwin			1.9008/2
GT	0.9700	0.9240	1.8940
VGGT	0.9730	0.9460	1.9190
VGGT- $\Omega$	0.9720	0.9275	1.8995
4DGS	0.9935	0.9065	1.9000
4C4D	0.9720	0.9195	1.8915

Appendix

Table 44: Event and action adherence score comparison of different world models.

Method	Event Editing	Subject Action	EAA
Wan			1.9140/2
GT+Instruction	0.9615	0.9420	1.9035
GT+Previous Prompt	0.9645	0.9515	1.9160
GT+Current Prompt	0.9690	0.9535	1.9225
Cosmos			1.9150/2
GT+Instruction	0.9760	0.9360	1.9120

Appendix

Table 45: Event and action adherence score comparison of different prompts.

Method	Physical Commonsense (PLA)
RoboTwin	0.6486/1
GT	0.6657
VGGT	0.6595
VGGT- $\Omega$	0.6687
4DGS	0.5863
4C4D	0.6630

Appendix

Table 46: Physical law adherence score comparison of different world models.

Method	Physical Commonsense (PLA)
Wan	0.6308/1
GT+Instruction	0.6131
GT+Previous Prompt	0.6331
GT+Current Prompt	0.6462
Cosmos	0.6483/1
GT+Instruction	0.6475

Appendix

Table 47: Physical law adherence score comparison of different prompts.

Method	BLEU	CLIP	Logics	VLM-1	Subject Adherence	Scene Adherence	SA
RoboTwin							4.8047/6
GT	1.0000	1.0000	1.0000	0.9532	0.9235	0.9460	5.8227
VGGT	0.4531	0.9316	0.9400	0.9383	0.9077	0.9160	5.0867
VGGT- $\Omega$	0.4679	0.9346	0.9450	0.9389	0.9123	0.9160	5.1147
4DGS	0.2048	0.8617	0.4925	0.8719	0.1315	0.8660	3.4284
4C4D	0.3324	0.9086	0.8200	0.9132	0.7010	0.8960	4.5712

Appendix

Table 48: Semantic adherence score comparison of different world models.

Method	BLEU	CLIP	Logics	VLM-1	Subject Adherence	Scene Adherence	SA
Wan							4.7986/6
GT+Instruction	0.2475	0.8957	0.7900	0.9127	0.7925	0.9440	4.5824
GT+Previous Prompt	0.2899	0.9025	0.8975	0.9193	0.9005	0.9720	4.8817
GT+Current Prompt	0.2952	0.9075	0.9175	0.9205	0.9170	0.9740	4.9317
Cosmos							4.9353/6
GT+Instruction	0.3571	0.9172	0.9050	0.9318	0.9207	0.8860	4.9178

Appendix

Table 49: Semantic adherence score comparison of different prompts.

Method	Normalized Score	Four-Level Score	IF
RoboTwin			1.2004/2
GT	0.8885	0.4367	1.3252
VGGT	0.8605	0.4100	1.2705
VGGT- $\Omega$	0.8745	0.4217	1.2962
4DGS	0.6225	0.3108	0.9333
4C4D	0.7900	0.3867	1.1767

Appendix

Table 50: Instruction following score comparison of different world models.

Method	Normalized Score	Four-Level Score	IF
Wan			1.4009/2
GT+Instruction	0.6800	0.5175	1.1975
GT+Previous Prompt	0.8245	0.5208	1.3453
GT+Current Prompt	0.9375	0.7225	1.6600
Cosmos			1.4045/2
GT+Instruction	0.8320	0.4192	1.2512

Appendix

Table 51: Instruction following score comparison of different prompts.

Method	Collision Count (CS)
RoboTwin	0.9385/1
GT	0.9384
VGGT	0.9388
VGGT- $\Omega$	0.9396
4DGS	0.9360
4C4D	0.9396

Appendix

Table 52: Collision safety score comparison of different world models.

Method	Collision Count (CS)
Wan	0.8762/1
GT+Instruction	0.8333
GT+Previous Prompt	0.8973
GT+Current Prompt	0.8981
Cosmos	0.9491/1
GT+Instruction	0.9435

Appendix

Table 53: Collision safety score comparison of different prompts.

Method	Action Planner	Data Engine	VLM-2	TS
RoboTwin				2.4340/3
GT	0.9825	0.5875	1.0000	2.5700
VGGT	0.9800	0.5750	0.9825	2.5375
VGGT- $\Omega$	0.9850	0.6000	0.9875	2.5725
4DGS	0.5475	0.3750	0.9875	1.9100
4C4D	0.9825	0.6125	0.9850	2.5800

Appendix

Table 54: Task success score comparison of different world models.

Method	Action Planner	Data Engine	VLM-2	TS
Wan				1.6692/3
GT+Instruction	0.3525	0.2075	0.9050	1.4650
GT+Previous Prompt	0.5150	0.3050	0.8375	1.6575
GT+Current Prompt	0.5850	0.3450	0.9550	1.8850
Cosmos				2.2025/3
GT+Instruction	0.7800	0.4550	0.9050	2.1400

Appendix

Table 55: Task success score comparison of different prompts.

World Model Evaluation

World Model Evaluation measures whether learned simulators predict consistent, physically plausible, action-responsive futures that help agents plan and act.

See all 212 papers

RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation
Apr 21, 2026Feng Jiang, Yang Chen, Kyle Xu +8Robotic Manipulation Robot Manipulation Benchmarks
Peking University · Tsinghua University · Lightwheel
TriWorldBench: A Tri-View Consistency Perspective on Embodied World Models
Sep 22, 2026Xuanyi Liu, Haofeng Wang, Ruiqi Li +9World Model Learning Robot Manipulation Benchmarks
Peking University · Tsinghua University · Beihang University +3
RoboWorM: Evaluating Movement Realism, Consistency and Acceleration Affects in Embodied Video World Models
Oct 24, 2026Doanh Le Thien, Viet-Thanh Nguyen, Pham Tri Quang +18World Model Evaluation Video World Models
Center for AI Research, VinUniversity · VinRobotics · MPI-INF +9

Apr 21, 2026J/K move · Enter open · S save

RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation

Feng Jiang, Yang Chen, Kyle Xu, Yuchen Liu+7

Peking University · Tsinghua University · Lightwheel

Robotic Manipulation Robot Manipulation Benchmarks World Model Evaluation Video World Models

Open paper AskPDF arXiv

Recent advances in large-scale video world models have enabled increasingly realistic future prediction, raising the prospect of using generated videos as scalable supervision for robot learning. However, for embodied manipulation, perceptual realism alone is not sufficient: generated interactions must also be physically consistent and executable by robotic agents. Existing benchmarks provide valuable assessments of visual quality and physical plausibility, but they do not systematically evaluate whether predicted behaviors can be translated into executable actions that complete manipulation tasks. We introduce RoboWM-Bench, a manipulation-centric benchmark for embodiment-grounded evaluation of video world models. RoboWM-Bench converts generated human-hand and robotic manipulation videos into embodied action sequences and validates them through execution in physically grounded simulation environments. Built on real-to-sim scene reconstruction and diverse manipulation tasks, RoboWM-Bench enables standardized, reproducible, and scalable evaluation of physical executability. Using RoboWM-Bench, we evaluate state-of-the-art video world models and observe that visual plausibility and embodied executability are not always aligned. Our analysis highlights several recurring factors that affect execution performance, including spatial reasoning, contact prediction, and non-physical geometric distortions, particularly in complex and long-horizon interactions. These findings provide a more fine-grained view of current model capabilities and underscore the value of embodiment-aware evaluation for guiding physically grounded world modeling in robotic manipulation.

Abstract

World Model Evaluation

Explore similar work

RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation

TriWorldBench: A Tri-View Consistency Perspective on Embodied World Models

RoboWorM: Evaluating Movement Realism, Consistency and Acceleration Affects in Embodied Video World Models