Awomo-SimDataEngine: Agentic Simulation-ReadyWorld Generation

Awomo-PhysicalRSI Team; Danjiao Ma; Enhui Ma; Haohan Liu; Heng Jia; Hui Shan; Jianhua Xu; Jiahuan Zhang

cs.ROOct 1, 2026

Awomo-SimDataEngine: Agentic Simulation-ReadyWorld Generation

Authors: Awomo-PhysicalRSI Team, Danjiao Ma, Enhui Ma, Haohan Liu, Heng Jia, Hui Shan, Jianhua Xu, Jiahuan Zhang,

Organizations: Awomo-PhysicalRSI Team

Abstract

Generating useful robot-training data requires more than visually plausiblescenes: objects must support interaction, placements must remain physicallyvalid, and tasks must admit repeatable execution. We presentAwomo-SimDataEngine, an agentic system that connects asset and scenegeneration to robot demonstration synthesis. Shared asset services providerigid and articulated objects, including structure-grounded part and jointgeneration with ISArt. Scene generation supports two complementary routes:Unravel reconstructs editable scenes from images, while SimForge buildssingle-room and multi-room environments from text. A graph-native harnesscoordinates construction, validation, andbounded repair, routing failures to the responsible module while retainingunaffected scene state. PolicyForge binds validated worlds to tasks and robotembodiments to produce replayable demonstrations. Evaluations cover assetgeometry, scene quality, and downstream policy learning. On MuJoCo-basedLIBERO-Plus, co-training with Isaac Sim demonstrations improves the overallsuccess rate of a World-Action Model (WAM) from $77.17\%$ to $89.43\%$ . Goal and spatialsuccess improve by $31.66$ and $6.25$ percentage points, respectively.These results support the utility of the generated data for cross-simulatorpolicy training, with more limited gains on long-horizon tasks.

Figures & tables

Figure 1: Awomo-SimDataEngine overview. (1) The graph-native harness coordinates execution, validation, and bounded repair. (2) The illustrated image-driven route uses Unravel to reconstruct scenes using acquired or generated assets, including articulated assets generated by ISArt. PolicyForge synthesizes demonstrations from validated worlds with task, robot, and reset specifications. (3) A schematic example illustrates failure detection and owner routing, followed by joint-origin correction and motion revalidation.

System	World construction	Connection to policy learning
EmbodiedGen V2	Simulation-ready assets, affordances, task-driven assembly, and editable worlds	Simulator deployment and embodied policy training
SimFoundry	Video-based digital twins and object, scene, and task cousins	Real-to-sim policy evaluation and sim-to-real learning
Awomo-SimDataEngine	ISArt-assisted assets, Unravel image reconstruction, and SimForge text-driven scenes, coordinated by a shared harness	PolicyForge demonstrations in Isaac Sim; measured co-training utility on MuJoCo benchmarks

Table 1: System positioning. Conditions, construction procedures, and evaluation settings differ; this comparison describes scope rather than a matched performance ranking.

Figure 2: ISArt articulated-asset generation. Structural planning supplies part and kinematic priors for learned geometry generation. Agentic binding constructs a URDF, and rendered simulation feedback guides parameter repair, structural re-planning, or geometry regeneration.

Figure 3: Unravel single-view scene generation. Verified removals expose objects and collect assets and point evidence. Reverse-order recomposition registers each asset to its attributed cluster, followed by support-aware settling and a final audit.

Method	CLIP-I $\uparrow$	LPIPS $\downarrow$	Edge IoU $\uparrow$
GPT-6 Astra (Ultra/ExtraHigh)	0.803	0.649	0.107
Unravel (w/ finetuned GPT-6)	0.854	0.632	0.156
$\Delta$ (rel.)	$+6.4\%$	$-2.6\%$	$+45.8\%$

Table 2: Scene-level image similarity against the reference photo. Mean over 20 scenes. Renders use the recovered photo viewpoint. Bold marks the better value per metric.

\Delta

is the relative change of Unravel over GPT-6 Astra, computed from the displayed means.

Room-level generation (179 prompts)
Method	#Obj	CNT $\uparrow$	ATR $\uparrow$	OOR $\uparrow$	OAR $\uparrow$	ACC $\uparrow$	NAV $\uparrow$	COL $\downarrow$	STB $\uparrow$	OOB $\downarrow$
HSM [ Pun et al., 2025 ]	$22.7\pm 2.6$	$60.6\pm 4.2$	$61.5\pm 7.3$	$30.6\pm 5.7$	$66.2\pm 6.6$	$88.9\pm 1.7$	$99.4\pm 0.4$	$20.6\pm 3.0$	$45.2\pm 3.9$	$5.5\pm 1.1$
Holodeck [ Yang et al., 2023 ]	$23.0\pm 1.4$	$44.2\pm 4.3$	$44.4\pm 7.2$	$16.8\pm 4.6$	$38.0\pm 7.1$	$84.9\pm 1.4$	$99.6\pm 0.2$	$12.3\pm 2.3$	$31.9\pm 2.7$	$0.8\pm 0.4$
SceneWeaver [ Yang et al., 2025a ]	$13.5\pm 1.0$	$41.8\pm 3.9$	$29.1\pm 6.7$	$15.3\pm 4.2$	$31.6\pm 6.5$	$77.5\pm 2.7$	$98.1\pm 1.2$	$12.5\pm 2.5$	$37.3\pm 4.4$	$0.0\pm 0.0$
I-Design [ Çelen et al., 2024 ]	$13.0\pm 1.0$	$69.3\pm 4.0$	$50.3\pm 7.2$	$28.6\pm 5.5$	$66.6\pm 6.7$	$70.1\pm 2.8$	$95.9\pm 1.4$	$3.0\pm 1.5$	$60.8\pm 4.6$	$4.3\pm 2.0$
LayoutVLM (Cur.) [ Sun et al., 2025 ]	$11.2\pm 1.6$	$41.0\pm 4.0$	$25.6\pm 6.6$	$14.1\pm 4.0$	$22.4\pm 6.1$	$93.8\pm 1.9$	$99.7\pm 0.2$	$25.9\pm 4.4$	$19.4\pm 3.6$	$6.2\pm 2.1$

Table 3: Text-to-scene quality under SceneEval and SceneSmith protocols [ Tam et al., 2025 ; Pfaff et al., 2026 ] . Room-level and house-level cohorts are kept separate. Values are reported means

\pm

95% confidence intervals where available. SAGE † uses a separate 30-scene cohort and native collision/stability definitions; it is a reference, not a matched ranking against the other rows [ Xia et al., 2026 ] . #Obj: object count; CNT: count agreement; ATR: attribute correctness; OOR: object–object relations; OAR: object–architecture relations; ACC: accessibility; NAV: navigability; COL: collision; STB: stability; OOB: out-of-bounds. SimForge is the text-driven scene module of Awomo-SimDataEngine.

Figure 4: Qualitative examples of PolicyForge robot interactions across different embodiments and scene configurations. The $4\times 4$ grid shows bedroom scenes in the top two rows and coffee-shop scenes in the bottom two rows. Examples illustrate variation in robot embodiment, object arrangement, and viewpoint.

Figure 5: PolicyForge pipeline. Given a scene, task description, robot embodiment, and target episode count $N$ , an LLM analyzes the task and decomposes it into atomic tasks when needed. A task manager selects suitable execution backends. Simulation execution in Isaac Sim produces synchronized observations, states, and actions, which are assembled into episodes. Episodes that pass task-success validation are exported in LeRobot format. The microwave images illustrate an ISArt input asset and its simulated manipulation.

Co-training data	Selected step	Success (%)	Gain (pp)
Baseline	9K	77.17	—
SimData-Easy	13K	78.76	+1.59
SimData-Lite	13K	78.27	+1.10
SimData-Hard	13K	81.06	+3.89
SimData-Ultra	13K	89.43	+12.26

Table 4: WAM success on LIBERO-Plus at each configuration’s best evaluated checkpoint up to and including 13K steps, selected by overall success. The gain is measured against benchmark-only training in percentage points. All configurations use the same 1,003-episode evaluation set.

Figure 6: Best-checkpoint co-training comparison. Best observed overall WAM success on LIBERO-Plus up to and including 13K steps. SimData-Ultra gives the highest score.

Figure 7: Generalization at the selected best-overall checkpoints. Each configuration is selected by its highest overall success up to and including 13K steps; all four suites use that selected checkpoint. Gains are largest for goals, followed by objects and spatial configurations; the long-horizon improvement is limited.

Figure 8: WAM learning curves on LIBERO-Plus. Points correspond to the reported evaluated checkpoints up to and including 13K steps. Ultra reaches

84.75\%

at 5K and its best score within this budget is

89.43\%

at 13K. Main comparisons select each configuration’s highest-overall-success checkpoint within the same budget.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Observation	Owner	Repair scope	Evidence
Incomplete or inconsistent object-removal plan or image edit	Unravel Planning / Editing	Affected removal round and image region	Plan, mask, and edit-verification evidence
Missing, rejected, or semantically wrong asset	Asset	Named request and dependent scene instances	Acquisition and semantic evidence
Mismatch between an asset and its observed point cluster	Registration	Named object and attributed point cluster	Point-cloud, mask, and alignment evidence
Residual penetration, physical instability, or simulator infeasibility	Physics / Settling	Affected object instances and local recomposition workset	Physical feasibility evidence
Validation regression without a hard geometry failure	Critique	Structured modification plan with a bounded attempt budget	Perceptual and validation evidence
Missing or invalid simulator package, frame, or strict-export predicate	Export / Packaging	Package manifest and coordinate/export artifacts; return upstream only when evidence identifies a source-geometry or physics cause	Packaging evidence

Appendix

Table 5: Owner-aware repair contract. The repair scope follows the causal owner when it can be identified; unrelated validated state remains protected by the checkpoint policy.

Group	Required meaning
Condition identity	Stable cohort identifier, modality label, source provenance, calibration status, and adapter/version owner.
Scene request	Rooms, architectural elements, object requests, attributes, dimensions, relations, and material/style constraints.
Task request	Goal, ordered subgoals, operable entities, interaction type, robot constraints, and success predicates.
Compilation evidence	Conflicts between modalities, unresolved requirements, confidence, and the rule used to resolve each conflict.
Construction requirements	Asset, articulation, support, clearance, containment, and export prerequisites derived from the task.
Output contract	Typed scene–task specification consumed by the ordered planning, acquisition, construction, validation, and export stages; an optional policy handoff attaches only after the Accepted World.

Appendix

Table 6: Conceptual condition and typed-specification contract.

Field group	Required audit meaning
Condition and configuration	Source identity, modality, adapter/version, model choices, seeds, strictness, and declared budgets.
Scene and layout	Current typed scene state, room topology, relation diagnostics, and the stage at which the snapshot was taken.
Acquisition	Asset requests, candidate outcomes, accepted sources, rejection reasons, and terminal failures keyed by request identity.
Placement and physics	Poses, support relations, collision and stability evidence, hard violations, and object removals or restorations.
Critique and evidence	Structured visual, geometric, physical, and export observations with severity, entities, and candidate owner.
Control and recovery	Selected graph edge, retry budget, checkpoint choice, rollback history, regression status, and next edge.

Appendix

Table 7: Conceptual Harness-state field groups (implementation name: PipelineState ).

Group	Required audit meaning
World identity	Stable scene identity, condition linkage, stage, iteration, and checkpoint provenance.
Topology	Rooms, boundaries, openings, architecture, object membership, and typed spatial relations.
Object instance	Unique instance identity, semantic role, placement role, room ownership, dimensions, pose, and measured geometry.
Asset evidence	Non-placeholder visual and collision representations, material/physical properties, source provenance, and validation status.
Support lineage	Parent instance, concrete support surface, local pose, footprint, clearance, and sibling compatibility for every supported child.
Articulation	Links, joint type, axis, origin, limits, and actuation metadata for movable objects.

Appendix

Table 8: Conceptual SceneState and object-instance contract.

Layer	Contents	Strictness role
Condition	Modality, source identity, adapter/version, task, seeds, and split	Immutable cohort key
Scene state	Rooms, objects, poses, relations, assets, joints, support lineage, and materials	Candidate world state
Execution trace	Transitions, observations, owner decisions, budgets, checkpoints, and rollbacks	Harness audit trail
Asset evidence	Candidate sources, geometry, collision, material, and validation outcomes	Acquisition gate
Validation evidence	Semantic, visual, geometric, physical, export, and degradation records	Hard/soft gate inputs
Simulator package	Coordinate contract, dependencies, load result, and reset metadata	Strict export gate

Appendix

Table 9: Conceptual organization of the evidence package.

Figure 9: ISArt’s structure-grounded multi-part DiT. E and C condition the pretrained geometry backbone while retaining visual conditioning.

Group	Required audit meaning
Request identity	Stable request, candidate, and instance identities that survive acquisition, placement, and export.
Semantic request	Description, object role, dimensions, attributes, style context, and articulation requirements.
Lifecycle	Planned, attempted, adopted, rejected, failed, and superseded states with terminal reason.
Provenance	Retrieval or generation source, query/prompt summary, backend/version, and cache status.
Geometry contract	Canonical orientation, scale, visual geometry, collision geometry, decomposition audit, and alignment evidence.
Material and physics	Material references, mass/friction properties, and simulator-readiness status.

Appendix

Table 10: Conceptual asset-registry contract.

Figure 10: Layer-wise deconstruction and point-cloud fusion in Unravel. The upper block shows planner outputs, removal targets, and edited images across six rounds. Red boxes mark removal targets; green boxes mark protected objects. The lower block shows aligned per-round reconstructions and cumulative fused point clouds.

Component	Required contents
World	Accepted scene state, strict validation summary, simulator package, coordinate contract, and provenance.
Task	Task family, goal, ordered subgoals, target entities, affordances, joint targets, and success predicate.
Robot adapter	Embodiment, sensors, observation/action spaces, control horizon, and reset assumptions.
Randomization	Scene, asset, reset, and rollout seeds plus declared domain-randomization policy.
Rollout feedback	Terminal status, first blocker, owner, evidence, and recovery decision returned to the Harness.

Appendix

Table 11: Conceptual world-to-policy handoff.

Group	Required meaning
Identity	World, task, policy, robot, simulator, cohort split, and checkpoint identifiers.
Randomness and budget	Scene/reset/rollout seeds, horizon, environment steps, and declared resource budget.
Outcome	Success predicate, terminal status, return/length/latency summaries, and infrastructure validity.
Failure attribution	First blocker, causal owner, affected entity or surface, secondary failures, and recovery evidence.

Appendix

Table 12: Conceptual rollout record.

Component	Configuration
Video DiT	30 layers, $d_{\text{hidden}}=3072$ , $d_{\text{ffn}}=14336$ , 24 heads ( $d_{\text{head}}=128$ )
Action DiT	30 layers, $d_{\text{hidden}}=1024$ , $d_{\text{ffn}}=4096$ , 24 heads ( $d_{\text{head}}=128$ )
Text Encoder	UMT5-XXL (frozen, precomputed embeddings, context_len = 128)
Input Cameras	2 cameras (front + wrist), $224\times 224$ each, concat $\rightarrow$ $224\times 448$
Video Context	33 frames, action_video_freq_ratio = 4
Action Space	7-DoF: 6 delta ( $xyz$ + $rpy$ ) + 1 absolute gripper ( $0$ =close, $1$ =open)

Appendix

Table 13: WAM architecture configuration.

Hyperparameter	LIBERO-Pro (1-node)	LIBERO-Plus (4-node)
GPUs	8 ( $1\times 8$ )	32 ( $4\times 8$ )
Batch size (per GPU)	16	24
Global batch size	128	768
Learning rate	$1\times 10^{-4}$	$2.449\times 10^{-4}$ ( $=\sqrt{6}\times 10^{-4}$ )
LR schedule	Cosine	Cosine
Epochs	10	10

Appendix

Table 14: Training hyperparameters. The LIBERO-Plus configuration uses

\sqrt{6}\times 10^{-4}\approx 2.449\times 10^{-4}

to compensate for the

6\times

larger global batch size.

Config	Data Description	Sampling Weight	Frame Ratio
Baseline	LIBERO-Plus only	—	—
+ SimData-Hard	Richer scenes with challenging tasks	0.308	1 : 1
+ SimData-Ultra	Multi-source mixture with robot initial-state and language augmentation	varied	1 : 1

Appendix

Table 15: Co-training data mixing for LIBERO-Plus experiments. LIBERO-Plus provides the benchmark demonstrations.

Benchmark	Simulator	Scale	Details
LIBERO-Plus	MuJoCo	1,003 episodes	4 suites (Obj/Spa/Goal/L-10), 8 GPU / 16 workers
LIBERO-Pro	MuJoCo	1,000 episodes	200 configs $\times$ 5 trials, 4 suites $\times$ 5 dimensions

Appendix

Table 16: Evaluation protocol summary for all benchmarks.

Method	CD $\downarrow$	[email protected] $\uparrow$	[email protected] $\uparrow$
PartCrafter	0.421	0.229	0.395
OmniPart	0.305	0.306	0.540
SPARK	0.278	0.314	0.505
TRELLIS.2 + P3-SAM + X-Part	0.268	0.431	0.627
ISArt	0.163	0.604	0.789

Appendix

Table 17: ISArt part geometry on the 60-object benchmark.

Method	AxisErr $\downarrow$	PivotErr $\downarrow$	TypeErr $\downarrow$	Success $\uparrow$
Articulate AnyMesh	0.72	0.02	46.67%	80%
Articulate-Anything	0.12	0.17	21.11%	90%
SPARK	0.38	0.19	23.61%	100%
ISArt	0.07	0.05	13.33%	100%

Appendix

Table 18: ISArt URDF binding with fixed ground-truth geometry.

Input	Image only	E only	C only	E+C
Input image (missing file)	a_B0.png (missing file)	a_E_only.png (missing file)	a_C_only.png (missing file)	a_EC.png (missing file)
(a) Microwave
Input image (missing file)	b_B0.png (missing file)	b_E_only.png (missing file)	b_C_only.png (missing file)	b_EC.png (missing file)
(b) Trash bin 1

Appendix

Figure 11: ISArt inference and training ablations for structural conditioning E and C. The examples support complementary effects; no aggregate ablation metric is reported.

	Masks		Assets (1–10)		Geometry
Method	Area ratio	UCR (%)	VLM	Human	Points (M)	UCR (%)
Direct	0.772	11.3	4.7	4.9	1.53	10.1
Unravel	1.000	0.0	8.2	8.6	2.81	0.0

Appendix

Table 19: Unravel deconstruction comparison. Mask area and under-coverage rates (UCR, coverage below 70%) apply to objects removed after the first round, using Unravel masks and clusters as references. Fidelity scores average all 160 objects; point counts average scenes.

Figure 12: Asset examples generated directly from the input masks (top) and from the best-observed deconstruction round (bottom).

Figure 13: Single-image scene comparisons. The shared-asset agentic variant receives Unravel’s meshes and masks, isolating pose and scale estimation.

Figure 14: Unravel reconstructions from six real indoor photographs. Each row shows the input and three reconstructed views.

Figure 15: Additional Unravel scenes with input images, reconstructed views, per-round edits, and extracted assets. These traces are qualitative examples.

Co-training Data	1K	3K	5K	7K	9K	11K	13K	Final
Baseline (no synthetic)	35.49	62.61	64.91	71.39	77.17	76.57	75.97	74.28
+ SimData-Easy	48.65	59.52	69.89	74.48	77.17	76.27	78.76	—
+ SimData-Lite	40.08	59.32	68.20	73.28	75.97	77.77	78.27	—
+ SimData-Hard	36.89	63.11	74.38	72.78	79.46	79.36	81.06	—
+ SimData-Ultra	65.90	81.95	84.75	88.14	86.94	87.64	89.43	89.63

Appendix

Table 20: LIBERO-Plus overall success rate (%) across training steps. Co-training models mix the corresponding additional data family at equal effective-frame ratio; the baseline uses benchmark data only. Evaluation: 1,003 episodes, 4 generalization suites. The final column uses 15K for the baseline and 14,580 for Ultra; final results for the other sources are not listed. Best per-step result in bold .

Co-training Data	Overall	Object	Spatial	Goal	LIBERO-10
Baseline (no synthetic)	75.97	84.13	81.25	54.83	84.52
+ SimData-Easy	78.76	86.51	85.83	58.30	85.32
+ SimData-Lite	78.27	85.71	83.33	61.78	82.94
+ SimData-Hard	81.06	87.70	91.25	64.09	82.14
+ SimData-Ultra	89.43	92.86	93.33	84.17	87.70

Appendix

Table 21: LIBERO-Plus per-suite success rate (%) at training step 13,000. All configurations evaluated at the same checkpoint step for fair comparison. Best per-suite result in bold .

Co-training Data	Step	Overall	Obj	Spa	Goal	L-10
Baseline	1K	35.49	55.16	41.67	24.32	21.43
	3K	62.61	84.52	71.25	37.07	58.73
	5K	64.91	79.37	72.08	39.77	69.44
	7K	71.39	83.73	81.25	51.74	69.84
	9K	77.17	84.52	87.08	52.51	85.71
	11K	76.57	80.16	85.42	61.39	80.16

Appendix

Table 22: LIBERO-Plus per-suite success rate (%) for all configurations across training steps. Suites: Object generalization (Obj), Spatial generalization (Spa), Goal generalization (Goal), Long-horizon sequential tasks (L-10).

Co-training Data	$\Delta$ Overall	$\Delta$ Object	$\Delta$ Spatial	$\Delta$ Goal	$\Delta$ LIBERO-10
+ SimData-Easy	+2.79	+2.38	+4.58	+3.47	+0.80
+ SimData-Lite	+2.30	+1.58	+2.08	+6.95	$-$ 1.58
+ SimData-Hard	+5.09	+3.57	+10.00	+9.26	$-$ 2.38
+ SimData-Ultra	+13.46	+8.73	+12.08	+29.34	+3.18

Appendix

Table 23: Improvement (

\Delta

pp) over Baseline at step 13,000 on LIBERO-Plus. Positive values indicate gains from co-training with the corresponding synthetic data.

Co-training Data	1K	5K	9K	13K	Final
Baseline (no synthetic)	13.8	40.6	45.3	45.9	46.7
+ SimData-Easy	15.8	42.2	47.7	47.6	47.9
+ SimData-Lite	14.6	40.8	46.6	46.8	46.9
+ SimData-Hard	13.1	43.3	45.7	46.9	47.1
+ SimData-Ultra	—	—	45.9	48.7	48.0

Appendix

Table 24: LIBERO-Pro overall success rate (%) across training steps. Evaluation: 200 task configurations

\times

5 trials = 1,000 episodes per checkpoint, spanning 4 base suites

\times

5 generalization dimensions. Baseline final step is 15,000; all co-training configs final step is 14,580. Best per-step result in bold .

Co-training Data	Object	Swap	Language	Task	Environ.
Baseline	79.50	3.00	93.50	18.00	39.50
+ SimData-Easy	77.00	4.00	93.50	21.00	44.00
+ SimData-Lite	77.00	3.00	90.50	18.50	45.50
+ SimData-Hard	80.50	4.00	96.00	18.50	36.50
+ SimData-Ultra	80.50	1.00	97.00	16.50	45.00

Appendix

Table 25: LIBERO-Pro per generalization dimension success rate (%) at the final training step. Five dimensions: Object (novel objects), Swap (position-swap), Language (semantic-language), Task (novel tasks), Environment (novel environments). 200 episodes per dimension per config. Best per-dimension result in bold .

Co-training Data	Overall	Object	Spatial	Goal	LIBERO-10
Baseline	46.70	47.60	55.20	44.00	40.00
+ SimData-Easy	47.90	46.80	57.60	44.40	42.80
+ SimData-Lite	46.90	48.80	53.60	46.00	39.20
+ SimData-Hard	47.10	48.80	57.60	45.60	36.40
+ SimData-Ultra	48.00	45.60	60.40	48.80	37.20

Appendix

Table 26: LIBERO-Pro per base suite success rate (%) at the final training step. Each suite contains 250 episodes (50 task configs

\times

5 trials). Best per-suite result in bold .

Co-training Data	Step	Overall	Obj	Spa	Goal	L-10
Baseline	1K	13.80	16.80	20.80	12.00	5.60
	5K	40.60	46.00	49.60	37.20	29.60
	9K	45.30	47.60	53.20	40.80	39.60
	13K	45.90	48.40	53.60	44.80	36.80
	15K	46.70	47.60	55.20	44.00	40.00
+ SimData-Easy	1K	15.80	24.00	25.20	12.80	1.20

Appendix

Table 27: LIBERO-Pro per base suite success rate (%) for all configurations across training steps. Each suite contains 250 episodes per checkpoint.

Co-training data	LIBERO-Plus (WAM)	LIBERO-Pro (WAM)
Baseline	77.17	46.7
SimData-Easy	78.76	47.9
SimData-Lite	78.27	46.9
SimData-Hard	81.06	47.1
SimData-Ultra	89.63	48.7

Appendix

Table 28: Best listed WAM checkpoint per configuration on LIBERO-Plus and LIBERO-Pro. Ultra’s LIBERO-Pro peak is at 13K, not its final checkpoint.

Synthetic Data Generation for Robotics

Synthetic data generation for robotics creates simulated scenes, sensor readings, and action trajectories to train and evaluate robots.

See all 288 papers

EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI
Jul 8, 2026Xinjie Wang, Liu Liu, Taojun Ding +93D Asset Generation Robotics Simulation
Horizon Robotics · WuwenAI
Scene2Demo: Self-Evolving Embodied Data Generation via Object-Action Graph
Feb 12, 2026Xiang Liu, Sen Cui, Guocai Yao +4Synthetic Data Generation for Robotics Robot Manipulation Datasets
SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation
Jun 26, 2026Nadun Ranawaka, Josiah Wong, Wei-Lin Pai +15Robotics Simulation Sim-to-Real Transfer
NVIDIA · Stanford University · The University of Texas at Austin +2

Jul 8, 2026J/K move · Enter open · S save

EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI

Xinjie Wang, Liu Liu, Taojun Ding, Andrew Choi+8

Horizon Robotics · WuwenAI

3D Asset Generation Robotics Simulation 3D Scene Generation Functional 3D Scene Generation Embodied AI

Open paper AskPDF arXiv

We present EmbodiedGen V2, a generative 3D world engine for building executable policy-ready environments for embodied intelligence. Sim-ready 3D asset generation has advanced rapidly, yet assembling such assets into policy-ready task environments remains largely manual, limiting scalable closed-loop learning. EmbodiedGen V2 addresses this gap through a unified sim-ready representation that connects cross-simulator assets, interaction affordances, task-driven worlds, large-scale multi-room scenes, and stateful Vibe Coding into a generative, editable, and reusable simulation pipeline. The generated environments support manipulation, navigation, mobile manipulation, cross-simulator deployment, and embodied policy training. In evaluation, the asset pipeline achieves 96.5% human acceptance and 98.6% collision success, and 83.3% of task-driven worlds are directly usable for downstream simulation without manual modification. Online reinforcement learning with generated environments further improves simulation success from 9.7% to 79.8%, and transfers to real robots with task success increasing from 21.7% to 75.0%. These results establish EmbodiedGen V2 as scalable simulation infrastructure for training, evaluating, and deploying embodied policies.

Abstract

Synthetic Data Generation for Robotics

Explore similar work

EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI

Scene2Demo: Self-Evolving Embodied Data Generation via Object-Action Graph

SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation