acceptodds
Under review as a conference paper at ICLR 2027

Efficient Learning of Mechanism Designs and Compositionality

Abstract

Learning mechanism behaviour requires a representation of how connected parts jointly constrain motion. We study whether exposing reusable constraint subassemblies improves prediction from limited data. Assur-KB represents planar mechanisms through irreducible modules, boundary interfaces, and solve dependencies while retaining source geometry. Our theoretical results characterize transfer through module interfaces and establish physical limits arising from missing interface information and accumulated sensitivity. To test the representation claim directly, the experimental design compares matched mechanism-curve pairs across training budgets within the same multilayer-perceptron family. This controlled comparison establishes the utility of the complete representation in the evaluated learning setting. Assur-derived features provide a geometric-mean advantage in sample efficiency over flattened adjacency and position inputs across budgets of 50-10,000 mechanisms. Separate experiments establish invariance under tested joint relabelings, higher valid-generation rates under dyadic composition, and a coarser structural index over 10.02 million mechanisms. In summary, the theory and experiments identify how physical constraint structure supports limited-data prediction and which interface conditions govern compositional reuse.

Primary area:
datasets and benchmarks
Keywords:
structured representations, agent design, mechanism synthesis, compositional generation, sample efficiency, datasets, benchmarks

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

This paper presents Assur-KB, a representation of planar linkage mechanisms organized into irreducible constraint modules, interfaces, and solve dependencies. Handcrafted features extracted from this decomposition feed a MLP that predicts a normalized trajectory. Additional experiments examine constructive sampling, structural indexing, and joint-relabeling invariance, while theoretical analyses discuss the conditions and limitations of reusing local predictors across different assemblies.

Critical strengths

Strength: The decomposition has a clear physical interpretation and explicitly identifies local constraint-solving dependencies. The dataset organization is useful, particularly its distinction between structural validity and geometric feasibility. It would be helpful for community research if it get released.

Critical weaknesses

Weakness: Unclear practical motivation. Classical kinematic solvers already compute these trajectories, and the paper itself describes efficient closed-form dyad solves. The main practical benefit of learning would be amortized computation, at the cost of approximation error and potential distribution-shift failures. However, the paper demonstrates neither an end-to-end speed advantage over traditional simulation nor a downstream inverse-design or large-scale optimization task showing that this trade-off is worthwhile.

Missing graph-based baselines. Holding the MLP family fixed provides a controlled architectural comparison, but flattened adjacency is a weak baseline for graph-structured mechanisms. Why not evaluate a GNN or graph transformer?

Representation benefits are confounded. The proposed input bundles decomposition with dimensionality changes, canonicalization, geometric normalization, and task-specific summaries. Equivalent preprocessing and input information are not clearly established for the baseline. Consequently, the gains cannot be attributed specifically to Assur structure. Comparisons against similarly normalized inputs and non-Assur geometric/topological features are needed.

Compositional learning is not directly demonstrated. The theory analyzes shared local predictors composed along a DAG, whereas the experiment uses a global MLP without explicit learned-module reuse. Joint-count-stratified splits also do not establish generalization to unseen topologies or deeper compositions.

Questions for authors
  1. How does Algorithm 1 detect and reject redundant constraints, particularly the three-boundary-joint example above?
  2. Why do structure-only features outperform the full vector, and why does removing branch information improve accuracy? What does this imply about the model’s use of geometry and assembly information?
Reviewer_NvLaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles the task of predicting the motion trajectories of planar linkage mechanisms. The authors propose a new representation based on breaking down mechanical constraints into modular steps. The approach first finds the smallest standalone mechanical building blocks (called Assur groups) and maps out their step-by-step solving order as a dependency graph (solve DAG). The mechanism is then turned into feature vectors that pack module features, assembly branches, and joint connections before being fed into a neural network. The authors train standard MLPs, and extensive experiments show clear gains in sample efficiency, valid design generation, and joint relabeling invariance.

Critical strengths
  1. Solid theoretical backing: The paper does not just propose a method; it backs it up with rigorous math, including proofs on error propagation, sensitivity blowup, and dimensional bounds.

  2. Thorough experiments: The evaluation is comprehensive, testing everything from small-data curve prediction and valid generation rates to clustering across 10 million mechanisms and joint permutation invariance.

  3. Clear writing: The paper is easy to read, with helpful figures that make the intuition behind turning physical constraints into ML features easy to grasp.

Critical weaknesses
  1. No evaluation on modern baselines: Skipping GNNs and Transformers just because they already have relational inductive biases is not convincing. The authors should test whether this representation actually helps modern SOTA models. If advanced models can figure out these topologies on their own and gain little from this representation, the practical impact of the work is fairly limited.

  2. Preprocessing overhead: Since the downstream model is just a lightweight MLP, the upfront time spent decomposing mechanisms and running rigidity checks is a real cost. The overall pipeline efficiency needs a fairer look.

Questions for authors

Since the advantage of the Assur representation fades once the dataset gets large enough, did you consider combining the Assur features with the standard adjacency baseline to maintain an edge across all data scales?

Reviewer_amoROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether making classical Assur decomposition explicit helps predict planar linkage output curves with limited training data. It contributes a corpus-scale representation, conditional results about prediction through module interfaces, and experiments on curve regression, constructive sampling, structural indexing, and relabeling invariance.

Critical strengths

The paper offers a useful way to describe mechanisms through small constraint-solving components and their connections. The physical examples are a particular strength: they explain why errors can grow across components and why retaining boundary information matters for predicting motion. The learning results suggest that the resulting features can help when training data are limited, and the paper appropriately reports that this advantage disappears at larger training sizes.

Critical weaknesses

There is a significant problem with the stated structural acceptance test. The definition requires independent constraints, but the algorithm checks rigidity and minimality without establishing independence. For example, one movable joint connected by three bars to three fixed points passes the written checks even though the constraints are redundant. This shows that the stated test is insufficient; whether it affects the stored dataset remains unknown. The authors need to show whether their implementation or proposed procedure prevents this problem and audit any potentially affected records.

The learning comparison also needs better controls. The proposed features include normalization, canonical ordering, geometric summaries, and structural counts, while equivalent preprocessing for the baseline is not documented. Using the same MLP does not by itself isolate the benefit of the decomposition. Matching these choices and testing simpler count-based features would make the result more informative. The reported 1.8 times sample-efficiency gain also needs an uncertainty estimate, particularly because the corrected tests do not establish improvements at the smallest training budgets. Essential training and data-split details should be supplied.

Section 5.1 identifies 88,910 matched pairs and stratification by joint count, but does not give their source/family composition, matching exclusions, or separation of repeated topology and related geometries across folds. This matters because the build’s structural diversity comes from Purwar Lab, whereas LINKS contains only dyads (Appendix E.3). The manuscript therefore does not establish whether the learning result extends beyond the source’s dyadic construction distribution. The target is explicitly normalized, arc-length-resampled path shape (section 2), which is a legitimate path-synthesis target. It does not retain input-angle timing, physical scale, or traversal correspondence, so it cannot by itself validate broader claims about kinematic behavior. I think the smallest useful correction is to document the cohort and split construction, report results by mechanism family and size, and add a topology-grouped evaluation if generalization to new assemblies is claimed.

Adding a small set of predicted-versus-true curves, with a worst-case geometric error measure, would also make the practical meaning of the Chamfer differences assessable to a new audience as linkage design is not a common topic in ICLR prior work and the readers may not know what the numbers signify

Section 3.4 and Appendix C omit essential training details like such as hidden widths, activations, optimization, training duration, model selection, feature preprocessing, seed handling, and the complete train/validation/test protocol. Table 4 also does not provide uncertainty for the interpolated sample-efficiency ratio. The source composition and selection of the 88,910 matched pairs are unspecified, as is the relationship to the 72,013 example maximum training pool. These omissions prevent reproduction and assessment of whether related geometries or topology families cross evaluation boundaries.

Questions for authors

See comments added above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.