acceptodds
Under review as a conference paper at ICLR 2027

USDCraft: Geometrically Grounded Programmatic Modeling of Articulated 3D Assets for Simulation

Abstract

Geometrically faithful and functional articulated 3D assets are essential for real-to-sim robot manipulation, where policies trained in simulation must transfer to physical objects. Recent mesh-based methods learn to infer articulation from annotated 3D assets, but deployment remains challenging when real-world objects fall outside the training distribution or their meshes are incomplete or corrupted. To address these limitations, we formulate articulated asset reconstruction as programmatic modeling grounded in partial geometric evidence and introduce USDCraft, a framework in which a pretrained LLM writes and revises executable programs for simulation-ready articulated assets without task-specific training. We propose source geometry analysis, which converts the source mesh into a metric textual description that distinguishes observed surface from unknown space, and iterative geometric rechecking, which re-encodes each candidate in the same representation so that discrepancies point to program edits while unobserved regions remain open to completion. Visual feedback and physical authoring guidance complete the modeling process, which produces articulated USD assets with explicit physical properties that load into Isaac Sim without manual adjustment. Experiments demonstrate state-of-the-art articulation recovery and validate USDCraft's effectiveness for real-to-sim-to-real robot manipulation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Articulated Object Reconstruction, Real-to-Sim, LLM Agents, Programmatic 3D Modeling, Simulation-Ready Assets

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QQcKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

USDCraft reconstructs simulation-ready articulated 3D assets from a source mesh and a reference image using a pretrained LLM without task-specific fine-tuning. The method converts source geometry into a metric textual representation that distinguishes observed surfaces from unknown space. An agent uses this representation, local mesh inspection, and iterative geometric and visual feedback to construct and revise an executable asset program. The output includes geometry, articulation, and physical parameters for simulation. The paper evaluates reconstruction on USDCraft-bench and the public Lightwheel dataset, studies text- and image-conditioned generation, demonstrates real-to-sim-to-real manipulation, and introduces the USDCraft-10k asset library.

Critical strengths
  1. Preserving measured geometry while allowing completion of unobserved regions is appropriate for imperfect meshes. The distinction between observed surface support and unknown space is useful. The executable representation also makes geometry, joints, and physical properties explicitly editable.
  2. Mini Workflow uses the same backbone and image/mesh inputs, providing an informative comparison. With Astra, USDCraft improves part F1 from 74.48% to 84.83% and reduces joint-axis angular error from 21.14 to 6.83 degrees (Table 3). These are substantial gains, and the repeated-run results in Table 9 provide additional support for their consistency.
  3. The experiments include the external Lightwheel benchmark with 243 objects across 14 categories, multiple model families, cumulative and removal ablations, and sensitivity analyses excluding particular sources of benchmark assets. This coverage strengthens the empirical case beyond the authors’ own generated examples.
  4. The robot experiments connect asset reconstruction to policy transfer rather than relying solely on visual quality. The paper also reports runtime and API costs, and discusses limitations involving fine structures and deformable objects, helping readers assess the method’s practical scope.
Critical weaknesses
  1. Section 3.2 and Appendix A.4 describe the sampling procedure and grid resolutions, and Appendix G.6 provides a useful encoding example. However, the algorithms for detecting and summarizing protrusions, openings, and observed interior structures are not described sufficiently. The relevant geometric criteria, thresholds, and treatment of noisy or incomplete surfaces are unclear. Since these summaries provide measurement anchors for reconstruction, this limits reproducibility of a central component.
  2. Tables 6 and 7 support the usefulness of source geometry analysis as a module, but this module combines axial slices, local numerical summaries, and on-demand mesh inspection. The complete system also uses images and subsequent geometric feedback. Consequently, the results do not isolate how much geometric information the fixed encoding preserves or how effectively the LLM uses it. The largest incremental improvement in Table 6 occurs when adding the authoring harness, making a more targeted analysis of the encoding particularly valuable.
  3. Although Lightwheel provides external validation, Table 2 does not include Mini Workflow, leaving the direct comparison against a general-purpose agent primarily on the authors’ benchmark.
Questions for authors
  1. Could you evaluate the same fixed encoding under three reconstruction settings: (a) rule-based reconstruction, (b) GPT reconstruction without an image, and (c) GPT reconstruction with the reference image? Additional source-mesh queries and source-based geometric feedback should be disabled, with the same encoding information available across settings and matched model, tools, and inference budgets between (b) and (c). Surface-based evaluation of the observed geometry would help distinguish encoding/decoding distortion, the model’s ability to use the representation, and the additional contribution of the image. Since the encoding is lossy, the target should be fidelity at the represented resolution; the rule-based result would serve as a reference, not a theoretical reconstruction upper bound.
  2. Could you report a Mini Workflow comparison on Lightwheel, or a representative subset, ideally with comparable inference budgets? This would help establish whether the gains over the general-purpose agent extend to the external benchmark and clarify the quality-cost trade-off.
3 comments
AuthorsOpenReview ↗

Response to Reviewer QQcK [1/2]

We thank the reviewer for the detailed and constructive review. The three points are addressed below and in the revised paper.

W1. Detection and summarization of protrusions, openings and interior structures

Because of the page limit, the submission kept this part brief. The revision states the criteria in Section 3.2 and lists the main criteria and thresholds in Appendix A.4 (Table 9). The detectors are deterministic and involve no learned model.

  • Protrusions. Each of the six axis-aligned faces of the source is resampled into a depth map at pitch h=max⁡(2 mm,ℓ/400)h=\max(2\,\text{mm}, \ell/400), where ℓ\ell is the longest extent of the source. A protrusion is a connected region that rises at least 1.5h1.5h above the face depth median-filtered over 60 mm and is at least 8 mm wide.
  • Recesses and openings. On the same depth maps, a connected region at least 6 mm below the reference depth of the face rim is a recess, and one at least 15 mm below it, or without any surface, is an opening. Both need an area of at least 4 cm² and a side of at least 15 mm.
  • Interior fittings. On slices of the 192-cell fine grid, these are surface islands inside the body outline that do not touch its walls. A lone island needs at least 4% of the body footprint and 25 × 25 mm.

The thresholds are physical lengths, chosen once during development and shared by all objects and benchmarks. Each detection is summarized as one line of text with its position, size and depth, without a semantic label, and Appendix G.6 now shows this output for a scanned toaster.

Noisy or incomplete surfaces. The source mesh is never repaired or filled, and unsampled cells stay unknown. Reference depths are medians or percentiles, so burrs and seams do not shift them, and detections on small or sparse fragments are dropped. Cavity depths are reported as lower bounds. Since a missing detection does not establish a flat face, the agent can remeasure any region with a mesh query that bypasses the detectors.

W2 and Q1. Fidelity of the fixed encoding

We ran the requested experiment on the 100 USDCraft-bench objects (Appendix D.3). Every setting starts from the same fixed encoding and has no further access to the source mesh.

  • (a) A rule-based decoder reconstructs the object from exactly the summaries the agent reads, without an LLM. As a reference, a second decoder voxelizes the full slice maps.
  • (b) USDCraft–Astra reconstructs from the encoding without the reference image.
  • (c) The same as (b) with the reference image.

In (b) and (c), source renders, source mesh queries and source-based geometric feedback are disabled. The agent keeps renders of its own candidate, which involve no source information. The two settings share the model, reasoning effort, tools and the budget of 80 turns, and differ only in the image. Outputs are compared with the observed source surface in the source frame. The slice threshold is the pitch of the slice grid, the resolution that the encoding represents, and the fine threshold is the pitch of the 192-cell grid. CD is the symmetric Chamfer distance in percent of the longest source extent. As the reviewer suggests, the rule-based rows are references and not upper bounds.

SettingSlice R ↑Slice P ↑Slice F ↑Fine F ↑CD ↓Part F1 ↑AE° ↓LE ↓
(a) Rule-based, agent's summaries67.683.472.432.02.72–––
Rule-based, full slice maps93.595.594.213.71.63–––
(b) USDCraft, encoding only94.995.294.852.60.9575.613.40.097
(c) USDCraft, encoding + image95.795.695.455.60.8779.410.80.064
Complete USDCraft98.396.697.277.40.5484.86.80.044

The table separates the three effects the reviewer asks about.

  • Distortion of the encoding. The encoding is lossy but keeps the observed surface at its own resolution. Voxelizing the slice maps recovers 93.5% of the source surface at the slice threshold, while its F-score drops to 13.7% at the fine threshold.
  • Use of the encoding by the model. In (b) the agent reaches a slice F-score of 94.8%, 22.4 points above the rule-based decoder that reads the same summaries. It also reaches 52.6% at the fine threshold, because the summaries carry measurements from the fine grid.
  • Contribution of the image. The image adds little observed geometry, 0.6 points at slice resolution and 3.0 points at the fine threshold. Its contribution lies in parts and motion, where part F1 rises from 75.6% to 79.4% and AE falls from 13.4° to 10.8°.

The remaining gap to the complete system comes from mesh inspection and iterative geometric rechecking, which the fixed encoding cannot replace.

AuthorsOpenReview ↗

Response to Reviewer QQcK [2/2]

W3 and Q2. Mini Workflow on Lightwheel

We ran Mini Workflow–Astra on the full Lightwheel benchmark, with the same image and mesh, the same model, the same low reasoning effort and the same cap of 80 turns as USDCraft–Astra. All objects completed.

MethodEvaluationF1 ↑Static gIoU ↑Static PC ↓mIoU ↑Artic. gIoU ↑Artic. PC ↓OC ↓AE° ↓LE ↓
Mini Workflow–Astrain the source frame18.8-0.3470.2580.088-0.3490.2940.08417.00.193
USDCraft–Astrain the source frame87.00.5900.0500.6830.5550.0880.01110.20.009
Mini Workflow–Astracentered and scale-normalized70.10.2350.1290.4770.2230.1820.01512.20.023
USDCraft–Astracentered and scale-normalized87.00.5800.0490.6740.5460.0870.01110.10.010

The two evaluations answer different questions. In the source frame, each output is scored where it is placed, which tells whether the asset can directly replace the input mesh. With centering and scale normalization, prediction and reference are each centered and scaled to unit size before scoring, which forgives a global offset and measures only part structure and articulation.

Mini Workflow does not keep its output in the frame of the input mesh, so its assets need an extra alignment step before they can replace the source, and its F1 in the source frame is 18.8%. USDCraft builds the asset in the measured frame of the source and scores almost the same under both evaluations. With the offset forgiven, USDCraft still leads by 16.9 F1 points, so the gain over the general-purpose agent extends to the external benchmark. This normalized row is the one added to Table 2 and Section 4.2.

Quality and cost. Mini Workflow costs 0.38 USD per object and USDCraft 0.91 USD (Appendix D.11). USDCraft thus stays below one dollar per object, while it reconstructs the source mesh in place and with clearly better geometry. Overall, we believe USDCraft is the more practical choice when an asset has to match a given object.

We hope this addresses the reviewer's concerns, and we welcome further discussion.

We sincerely thank you for your constructive feedback. As the discussion period is closing soon, we would like to kindly ask whether our response has addressed your questions, and we are happy to discuss further.

Thank you for your time and consideration.

Reviewer_TPyhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces USDCraft, an agentic framework that utilizes LLMs to generate articulated USD assets through programmatic modelling. The authors also introduced procedures named "source geometry analysis" and "iterative geometric rechecking", in the hope of making LLMs better understand a geometry.

Critical strengths
  1. The proposed source geometry analysis offered an interesting perspective on helping LLMs understanding the 3D geometry.
  2. The results presented in the paper seem impressive.
Critical weaknesses

One of my major concerns is regarding the clarify of the method section. I find it quite difficult to follow and missing details. Below I give two examples.

  1. In section 3.2, despite my acknowledgement on the proposed analysis approach in the strengths section. There is not enough detail that allows full reproduction of the proposed method. For example, line 182 and 183 mentioned terminologies like "automatically detected features" and "mesh inspection", yet there is no explanations on what they are.
  2. In section 3.3, while the programmatic program representation is listed as the first contribution of this paper, there is virtually no detailed descriptions or examples on this representation.
Questions for authors

Please refer to the weaknesses, I would deeply appreciate a clarified version of the method section.

8 comments
AuthorsOpenReview ↗

Response to Reviewer TPyh [1/2]

We thank the reviewer for the careful reading. Because of the page limit, the submission kept the method section brief and left part of the implementation detail to Appendix A.4 and G.6. In the revision we bring these details into the main text and complete them, across Sections 3.2 to 3.4 and not only at the two places mentioned. The additions are shown in blue in the revised paper, and we summarize them below.

W1. Section 3.2, "automatically detected features" and "mesh inspection"

Automatically detected features. These are protrusions, openings and interior fittings that deterministic detectors find on the source mesh and report to the agent as metric text. No learned model is involved.

  • Protrusions and openings are found on depth maps of the six axis-aligned faces of the source. Let df(u)d_f(u) be the outward depth of the first surface at location uu on face ff, sampled at pitch h=max⁡(2 mm,ℓ/400)h=\max(2\text{ mm}, \ell/400) for a source of longest extent ℓ\ell, and let d~f(u)\tilde d_f(u) be its median over a 60 mm window. A protrusion is a connected region with df(u)−d~f(u)≥1.5hd_f(u)-\tilde d_f(u)\ge 1.5h that is at least 8 mm wide. An opening is a connected region that lies at least 15 mm below the reference depth rfr_f of the face rim, or that has no surface at all. A region sunk by at least 6 mm is reported as a recess.
  • Interior fittings are surface islands that stand free inside a slice, without touching the body walls.
  • The source mesh is never repaired or filled. Unsampled cells stay unknown, reference depths are medians or percentiles so that burrs and seams do not shift them, detections on small or sparse fragments are dropped, and cavity depths are reported as lower bounds.

The thresholds are physical lengths, chosen once during development and shared by all objects and benchmarks. Each feature is reported with its position, size and depth and carries no semantic label, so its interpretation is left to the agent.

Mesh inspection. This refers to two tools that the agent calls on demand to look at the source more closely than the fixed encoding allows.

ToolInputOutput
Source renderviewpointrendered view of the source mesh
Source measurementoptional query boxbounds and percentiles of the source vertices in the box, sampled points in three orthographic projections, per-component bounds, and the surface orientation from a plane fit

These definitions are now in Section 3.2. Appendix A.4 lists the main criteria and thresholds (Table 9) and all tools (Table 10), and Appendix G.6 shows the detector output for a scanned toaster.

AuthorsOpenReview ↗

Response to Reviewer TPyh [2/2]

W2. Section 3.3, description and example of the programmatic representation

An asset is a Python program that builds an ArticulatedObject. Each part is declared with its geometry, materials and mass properties, and each joint with its type, parent and child parts, origin, axis, motion limits and PhysX parameters. Compiling the program produces the USD asset. Section 3.3 now includes an abridged program reconstructed from the scanned toaster (Figure 3). A shortened excerpt is

def build_object_model():
  m = ArticulatedObject(name='two_slice_toaster')
  m.physics_material('metal_contact', static_friction=.25, dynamic_friction=.18)
  steel = m.material('stainless', rgba=(.35, .38, .38, 1), metallic=.85)
  # Housing as a hollow CadQuery casing
  body = m.part('housing', inertial=...)
  shell = rounded(.143, .233, .168, .027)
  body.visual(mesh_from_cadquery(shell.cut(rounded(.133, .223, .153, .023)), 'shell'))
  # Slots at the detected top-face openings
  ...
  elevator = m.part('elevator', inertial=..., rigid_body=RigidBodyProperties(physics_material='metal_contact'))
  # Lever and bread supports slide together
  m.articulation('bread_lowering', ArticulationType.PRISMATIC,
      parent=body, child=elevator, origin=Origin(xyz=(-.003, -.107, .124)), axis=(0, 0, 1),
      motion_limits=MotionLimits(lower=-.048, upper=0, effort=25, velocity=.15),
      physx_joint_resistance=PhysxJointFriction.stay_put(elevator_mass * 9.81))
  ...
  return m

The example also shows how the two sources of evidence meet in one program. The two bread slots follow the openings detected on the top face, while the heater chambers and the bread elevator are not visible in the scan and are completed from the object's function. Appendix A.2 lists the main calls and units (Table 8).

Q. A clarified version of the method section

The revised paper is uploaded, and we kindly invite the reviewer to read Section 3 there. Beyond the two points above, the revision also adds

  • the grid resolution of the axial encoding (Section 3.2),
  • what each geometric recheck reports and when the loop ends (Section 3.4),
  • the inputs and outputs of every inspection and recheck tool (Appendix A.4, Table 10),
  • the main program calls and units (Appendix A.2, Table 8).

The method itself is unchanged. We will also release the code, including the feature detectors and the authoring toolkit, to support reproduction. We hope the revised section is easy to follow, and we welcome further questions.

I thank the authors for their response. A part of my concerns are addressed. Yet I remain curious about the representational capability of the proposed program representation. Here are a few details I noticed in the paper's figures but not fully revealed by the code representation (here figures refer to those in the revision):

  1. In Figure 3, the shape description of "lever and bread supports".
  2. In Figure 3, the tiny slot/opening on the toaster that allows the lever to slide.
  3. In Figure 13, the "decorations" (or visual textures) on the reconstructed toaster and the box. If possible, I would encourage the authors to provide one or two complete programs for sufficiently complex objects to showcase the proposed method's representational capability, as well as helping with better understanding of the proposed approach.

I also read the added experiments in Appendix D.1, and did not seem to find a clear statement on whether USDCraft is ran in reconstruction or generation mode. It would be helpful to add a clarification on that.

Reviewer_GdpSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents USDCraft, a framework for generating and reconstructing simulation-ready articulated assets. Given a text prompt, a reference image or a source mesh, USDCraft uses a pretrained LLM to write, compile and refine executable asset programs without requiring task-specific training. Experimental results that evaluate geometric reconstruction and articulation recovery on two benchmarks show that USDCraft outperforms the other baselines and achieves high performance in robot manipulation tasks.

Critical strengths
  1. The paper is well written, has clear structure and presents useful results.
  2. The authors provide a thorough ablation in Tables 6 and 7 where they show how much each component contributes to the final system.
  3. In Sec 4.4 the authors provide strong results where they show that USDCraft generated assets are effective in sim-to-real transfer. It is significant that in the sim-to-real transfer experiments USDCraft loses only at most 10 points on the tasks mentioned in Table 4 while the other baselines lose 60-70 points.
  4. The geometric reconstruction results in Tables 1 and 2 and Figure 3 show that USDCraft outperforms the other baselines in most of the metrics.
Critical weaknesses
  1. Could the authors provide comparisons of USDCraft with more models such as PartCrafter [1], OmniPart [2], Spark [3] and SINGAPO[4]?

[1] PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers, Neurips 2025

[2] OmniPart: Part-Aware 3D Generation with Semantic Decoupling and Structural Cohesion, SIGGRAPH Asia 2025

[3] SPARK: Sim-ready Part-level Articulated Reconstruction with VLM Knowledge, CVPR 2026

[4] SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects, ICLR 2025

  1. How does USDCraft perform against procedural generated assets such as the ones mentioned in Infinigen-Articulated [5]. Could the authors state what are the benefits or limitations of USDCraft over procedural generation in sim-to-real experiments?

[5] Procedural Generation of Articulated Simulation-Ready Assets, Abhishek Joshi et. al

Questions for authors

Please see the questions mentioned in the Critical Weaknesses section. Apart from these:

  1. It would be useful to have some statistics about the USDCraft-10k library. For example, how many assets there are in each super category.
5 comments
AuthorsOpenReview ↗

Response to Reviewer GdpS [1/2]

We thank the reviewer for the positive assessment and the helpful suggestions. All three points are now covered in the revised paper.

W1. Comparison with PartCrafter, OmniPart, SPARK and SINGAPO

We ran PartCrafter [1], OmniPart [2], SPARK [3] and SINGAPO [4] on both benchmarks. Each method receives the reference image given to USDCraft and runs its official code with default settings. SINGAPO and SPARK query a vision-language model, and we replace their default model with GPT-6 Astra, the model used by USDCraft–Astra. PartCrafter (*) is also given the ground-truth part count. The metrics follow Tables 1 and 2. Since F1 counts a part as matched only when its joint type is correct, we also report Decomp. F1 on USDCraft-bench, which matches parts by geometry alone.

USDCraft-bench (100 objects, protocol of Table 1)

MethodDecomp. F1 ↑F1 ↑Static gIoU ↑Static PC ↓mIoU ↑Artic. gIoU ↑Artic. PC ↓OC ↓AE° ↓LE ↓
PartCrafter*21.70.0-0.2710.1530.116-0.2740.2410.059––
OmniPart35.70.0-0.2100.2070.179-0.2160.2740.035––
SINGAPO38.229.0-0.0300.1690.262-0.0390.2320.04757.70.317
SPARK41.19.7-0.0410.1240.250-0.0470.2810.03533.20.397
Articraft–Astra73.164.00.2040.0980.4220.1960.1730.01915.10.092
USDCraft–Astra91.184.80.6940.0520.7840.6700.0930.0106.80.044

Lightwheel (metrics of Table 2, prediction and reference independently centered and scale-normalized)

MethodF1 ↑Static gIoU ↑Static PC ↓mIoU ↑Artic. gIoU ↑Artic. PC ↓OC ↓AE° ↓LE ↓
PartCrafter*24.2-0.3400.1360.105-0.3420.2140.031––
OmniPart32.4-0.2400.2220.161-0.2440.2880.025––
SINGAPO25.9-0.1600.2630.242-0.1670.3290.05816.10.086
SPARK34.9-0.1900.1550.188-0.1970.3050.03115.90.268
USDCraft–Astra87.00.5800.0490.6740.5460.0870.01110.10.010

USDCraft–Astra leads the four methods on every metric on both benchmarks. All four are image-conditioned generation methods. They cannot reconstruct a given object from its mesh, and their outputs are limited by what the generative model has learned, which shows in both part decomposition and joint estimation. These results are in Appendix D.1 with a qualitative comparison, and SINGAPO and SPARK were added to Tables 1 and 2.

[1] PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers, NeurIPS 2025

[2] OmniPart: Part-Aware 3D Generation with Semantic Decoupling and Structural Cohesion, SIGGRAPH Asia 2025

[3] SPARK: Sim-ready Part-level Articulated Reconstruction with VLM Knowledge, CVPR 2026

[4] SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects, ICLR 2025

AuthorsOpenReview ↗

Response to Reviewer GdpS [2/2]

W2. Comparison with procedural generation (Infinigen-Articulated)

Infinigen-Articulated [5] provides hand-written procedural generators for 18 categories, and each generator samples random instances of its category without an image, mesh or text input. It therefore cannot reconstruct a given object, so we compare in the generation setting. For each of the 18 categories we write two text requests, generate one asset per request with USDCraft–Astra and Articraft–Astra, and take the first two public Infinigen-Articulated instances of the category. All assets are scored with the protocol of Table 5 (scores out of 5).

MethodConditionStructureArticulationCompletenessAppearance
Infinigen-Articulated–3.423.723.833.22
Articraft–Astra3.364.194.284.283.06
USDCraft–Astra3.864.284.394.443.69

Since Infinigen-Articulated cannot follow a request, it is scored against a category-level request that names only the object and asks for its typical parts and motions, and condition adherence does not apply. USDCraft–Astra is higher on every shared dimension.

Each Infinigen-Articulated generator is written by hand for one category. The table below gives the share of object categories in each dataset for which such a generator exists. USDCraft covers all of them without category-specific code.

MethodLightwheel (14 categories)USDCraft-10k (1,000 categories)
Infinigen-Articulated6 (42.9%)24 (2.4%)

Benefits and limitations for sim-to-real.

  • Benefit. A procedural generator samples random instances of a category, so a policy trained on its assets meets geometry that differs from the real object. USDCraft reconstructs the deployed object from a scan and an image with its measured dimensions and part placement, and policies trained on these assets transfer with a drop of at most 10 points (Section 4.4). USDCraft also covers new categories without category-specific code.
  • Limitation. Every USDCraft asset requires LLM inference, whereas a procedural generator samples further instances at no model cost once it is written.

Overall, USDCraft–Astra costs less than one dollar per asset (Appendix D.11), and in return it reproduces the deployed object and is not limited to a fixed set of categories. We therefore find it the more practical choice for sim-to-real transfer.

This comparison and discussion are in Appendix D.2, and Infinigen-Articulated is now discussed in related work.

[5] Procedural Generation of Articulated Simulation-Ready Assets, Abhishek Joshi et al.

Q1. Statistics of USDCraft-10k

USDCraft-10k contains 10,000 assets in 1,000 object categories, ten per category, grouped into six super categories and 30 scene types.

Super categoryCategoriesAssetsArticulatedStatic or floatingExample articulated categories
Culinary1651,650920730microwave oven, dishwasher
Household2382,3801,650730wardrobe, washing machine
Commercial1441,4401,040400filing cabinet, shopping cart
Industrial2552,5501,880670drill press, pallet jack
Medical5252044080hospital bed, benchtop centrifuge
Outdoor1461,460970490lawn mower, bicycle
Total1,00010,0006,9003,100

In total, 69% of the assets are articulated with revolute, continuous or prismatic joints. By physical scale, 3,080 assets are handheld, 2,200 tabletop, 3,970 furniture-sized and 750 large, such as industrial machines or vehicles. These statistics are in Section 4.5 (Figure 6) and Appendix D.8.

We hope this addresses the reviewer's questions, and we welcome further discussion.

We sincerely thank you for your constructive feedback. As the discussion period is closing soon, we would like to kindly ask whether our response has addressed your questions, and we are happy to discuss further.

Thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We have uploaded a revised paper that adds short clarifications and a few supplementary comparisons, shown in blue.

Method (Section 3)

  • Section 3.2 adds a short description of how protrusions and openings are detected and how incomplete surfaces are handled. The main criteria and thresholds are in Appendix A.4.
  • Section 3.3 adds a program example (Figure 3). The main program calls are listed in Appendix A.2.
  • Section 3.4 adds what each geometric recheck reports and when the loop ends.

Experiments (Section 4)

  • Tables 1 and 2 add SINGAPO and SPARK, and Table 2 also adds Mini Workflow on Lightwheel.
  • Section 4.5 adds the category statistics of USDCraft-10k (Figure 6).
  • Appendix D adds further image-conditioned baselines (D.1), a comparison with procedural generation (D.2), and an analysis of the fixed encoding (D.3).

Minor wording edits made for space are not colored. We hope the revision resolves the reviewers' main concerns, and we welcome further discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.