acceptodds
Under review as a conference paper at ICLR 2027

SourcePhys: Source-Conditional Physical Completion for Simulation-Ready 3D Assets

Abstract

Large 3D collections contain assets in very different states of physical readiness. Generated meshes may lack physical structure, while native simulation assets may already provide bodies, joints, and valid parameters. We present SourcePhys, a framework for source-conditional physical completion that preserves available evidence and infers unresolved attributes. It selects a part representation from the source, combines constrained vision-language inference with material priors and component geometry, and arbitrates direct and geometry-based mass estimates when they disagree. The completed physical records retain the origin of inferred values and can be exported to OpenUSD for simulation. We evaluate 775 assets from four source cohorts. On 150 randomly sampled Amazon Berkeley Objects, the full method reaches 43.6% median mass error; 743 assets have an original or repaired variant that passes cohort-specific OpenUSD and Isaac Sim checks. Controlled behavior probes and robot interactions further illustrate how the completed assets enter interactive workflows. These results offer a practical route from heterogeneous 3D content to physically grounded assets for embodied AI.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D assets, physical property estimation, source-conditional completion, vision-language models, simulation-ready assets, OpenUSD, embodied AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZRTdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SourcePhys, a practical pipeline for converting heterogeneous 3D assets into OpenUSD assets for simulation. Its main idea is to preserve existing source physics when available, otherwise reuse source structure or recover parts visually, and fill missing physical attributes with VLM predictions, engineering priors, geometry, and standard mechanics. The strongest quantitative component is mass estimation, evaluated on 150 ABO assets. The remaining Isaac Sim evaluation is much weaker: it mainly checks coarse executability such as loading, physics stepping, settling, and actuation, while the robot interaction study consists of only ten pick-and-place cases with one recorded run each and no benchmark comparison. Overall, the paper is useful as an engineering recipe for preparing assets for OpenUSD/Isaac Sim, but the methodological novelty and evidence for physical fidelity are limited.

Critical strengths

1.The paper provides a clear practical decomposition of what is required to turn heterogeneous 3D assets into executable OpenUSD assets. The distinction between source-authored physics, mesh/group structure, and visual-only inputs is useful, and the explicit routing logic makes the asset-processing problem concrete. The provenance-aware OpenUSD authoring is also well motivated from an engineering and reproducibility perspective.

2.The mass-estimation component is reasonably well motivated and quantitatively evaluated. The paper combines structural class, geometry, material priors, and a direct VLM estimate, and includes several ablations that clarify where the gain comes from. In particular, the structure-conditioned volume estimate provides a substantial improvement over treating all parts as solid.

Critical weaknesses

1.The evaluation does not convincingly establish that the completed physical properties are actually useful or physically meaningful in simulation. Beyond mass, properties such as friction, restitution, COM, inertia, and collision approximations are mainly assessed through executability or weak behavioral sanity checks. The large-scale evaluation primarily establishes that assets can be loaded, stepped, settled, or actuated, while the controlled behavior study uses only ten native assets and one execution per condition. For many of these properties, direct ground truth may indeed be unavailable; however, this makes a stronger simulation-based evaluation even more important. The paper should demonstrate that the completed properties lead to meaningfully better or more plausible behavior than simpler/default physics, rather than only showing that the assets execute.

2.The pipeline can accumulate substantial upstream errors, but this propagation is not analyzed. Many downstream quantities depend on a sequence of uncertain decisions: part decomposition, structure class, material hypothesis, density/volume estimation, part mass, and finally COM/inertia. In addition, the source-preserving routing assumes that existing rigid-body or object/material group structure is preferable when deemed “valid” or “usable,” but the paper does not provide a clear mechanism for determining whether a source partition is physically meaningful rather than merely syntactically valid or convenient for modeling/rendering. Errors at this early stage can systematically affect later estimates. The current experiments do not quantify the sensitivity of final simulation behavior to these upstream choices.

3.The heterogeneous-source framing is practically useful, but the methodological novelty of the overall framework is limited and difficult to isolate. Different cohorts are handled with substantially different procedures: native assets mostly preserve existing physics, grouped assets reuse source structure, and monolithic assets rely on a standard visual part-recovery pipeline. Other physical properties are largely filled using engineering priors and standard mechanics, with mass estimation being the most developed inference component. Overall, the paper feels more like a practical integration framework than a broadly novel method for physical completion. The highly heterogeneous datasets and evaluation protocols also make it difficult to isolate the contribution of the source-conditional routing itself.

Questions for authors

1.Can the authors provide a simulation-based evaluation that measures the utility of the inferred physical properties, rather than primarily executability? For example, how does full SourcePhys compare with default/simplified physics on tasks specifically sensitive to friction, restitution, COM, inertia, or collision geometry? 2.Can the authors provide downstream ablations such as default physics → +mass → +contact parameters → +COM/inertia → full SourcePhys, and evaluate how each stage changes behavior or task performance? 3.How sensitive are the final simulation behaviors to errors in part decomposition, structural class, material prediction, scale, and mass? A controlled perturbation or robustness study would help clarify whether errors compound significantly in the final OpenUSD asset. 4.How is a source hierarchy or object/material grouping determined to be “valid” or “usable”? How does the method detect source partitions that are syntactically well formed but physically inappropriate for defining rigid parts? 5.How does the method handle parts that are intrinsically multi-material or internally heterogeneous, rather than merely uncertain between several material labels? Does averaging material priors risk producing unrealistic effective density, friction, or inertia when different materials occupy distinct spatial regions? 6.What should readers understand by “simulation-ready” in this paper: primarily executable in Isaac Sim, or sufficiently realistic for downstream embodied tasks? The current evidence seems much stronger for the former than the latter.

Reviewer_MB9cOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper propose SourcePhys, a method that adds physical properties to 3D assets that come from various sources, including generated mesh, ABO mesh retrieval, and native OpenUSD). The core idea is about "source-conditional completion", where if the source already have valid physics properties such as bodies, joints, collisions, and mass, the method would keep it, and only complete the missing part. For missing attributes, a VLM predict category, structure class (solid, shell, thin plate, frame), material and scale, then material prior table gives density and friction. Mass is computed from structure-conditioned component volume times density. Moreover, there is a guard design that choose between geometry mass or VLM mass. Output is exported as OpenUSD. On validation, full method get 43.6% median APE, compared with 49.0% for Direct-VLM and 46.5% for Geometry+material.

Critical strengths
  • The idea of keeping valid sources while filling in missing fields reflects current issues with the method that tries to add physics to 3D assets, which might be useful for people building embodied AI data pipelines.
  • Table 1 results show that forcing all components to be solid gives a higher APE compared to the structure class variants, suggesting structure-conditioned volume offers benefits for physical validity.
Critical weaknesses

Overall, the paper's core contribution seems to be engineering integration rather than a new scientific discovery or technical idea.

As I look at each part of the method, most components are existing tools or common heuristics. For example, the “source-conditional completion” in Eq. 1 is basically “if the source value is valid, then keep it; otherwise, fill with the prediction,” and Eq. 2 is an if-else rule for choosing a route. The part discovery uses Blender extraction and SAM3 multi-view voting; the semantic inference uses an off-the-shelf VLM with a constrained JSON prompt; Eq. 3–4 are just the mean and variance of a mixture; and the volume formulas in Eq. 5 are textbook approximations with many hand-set constants. The disagreement guard is also four hand-written rules with fixed thresholds.

I understand that leveraging pre-trained foundation models or task-specific expert models is a common practice in many papers on physical identification of 3D assets. However, this paper's only finding is that hollow or thin objects should not be treated as solid when computing mass, which is intuitive and already known in practice.

Moreover, the related work section says this work is “complementary” to PhysX-Anything, PhysForge, UniPhysGen, and EmbodiedGen, but without any comparison, it is hard to understand what scientific insight it offers beyond these systems.

Finally, for the embodied experiment, Isaac Sim checks only show that the assets can run, not that the physics is correct: friction, restitution, center of mass, and inertia are never compared with any reference. There is no baseline with default or incorrect physics to compare against. Therefore, it's unclear whether the completed physics actually makes a difference in downstream embodiment.

Questions for authors

Here are the questions I have:

  • Beyond the engineering integration, what is the main generalizable technical idea or finding of this work?
  • Could you compare it with any of the complementary works you mentioned in the related work section?
  • For the hand-set rules and constants, were the guard thresholds and the volume constants chosen before evaluating on the 150 ABO assets?
  • Missing physical correctness beyond mass, as friction, restitution, center of mass, and inertia are never compared with a reference.
  • Can you compare robot task performance between assets from the proposed method and the same assets with default or incorrect physics?
Reviewer_gVdfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addressed heterogeneous 3D asset conversion for simulation-ready physical assets. The paper proposes source-conditional physical completion, which processes assets through frontends and performs constrained VLM inference over aligned views. The authors introduce geometry-conditioned derivation and a deterministic disagreement guard. The method was evaluated on 775 assets, demonstrating that source-conditional completion provides a practical route from heterogeneous 3D content to executable physical asset collections.

Critical strengths

This paper presents source-conditional completion, in which an evidence precedence rule preserves valid source-authored physics and selects part evidence for estimating missing fields across heterogeneous inputs. The structure-conditioned geometry and material priors provide estimates, and the deterministic guard arbitrates with a direct VLM estimate. This work also provides OpenUSD authoring that records the origin and fallback of inferred values.

Critical weaknesses
  1. Table 1 reports mass estimation on 150 ABO assets. Direct VLM tends to have close performances to the proposed methods. This work addresses the Disagreement Guard and Geometry-Conditioned Physical Derivation, but the performance gain appears to be limited. It would be helpful to discuss the isolated contributions of Geometry-Conditioned volumes and the guard when conditioned on source regime or object category?
  2. The disagreement guard requires further validation regarding some parameters, such as geometry-risk, thin/flexible-part evidence, or unconditional threshold.
  3. Material and structural priors, such as those in the engineering prior table, lack validation. Comparisons with measured material properties or an ablation of prior quality would be helpful.
  4. Behavior probes use ten assets with one run per condition (Table 3). The robot cases cover ten cases (Table 4). Whether these cases, with their limited scale, can serve as evidence of executability remains unclear.
  5. The VLM component requires further description regarding prompts and the effect of VLM errors on the method.
Questions for authors

See weakness. The paper presents a pipeline that preserves valid source-authored physics while completing only missing fields. However, the gains achieved by the disagreement guard tend to be limited. The practical reliability of the median APE accuracy requires discussion. Further discussion or description is required regarding a sensitivity analysis of the guard, an ablation of the material priors, and VLM component error propagation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.