acceptodds
Withdrawn from ICLR 2027

: MECHANISM-AWARE REFLECTIVE CORRECTION FOR ARTICULATED OBJECT CODE GENERATION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating functional articulated objects requires geometry, connectivity, and motion to agree. This is particularly challenging for mechanisms with coupled motion and closed kinematic chains: plausible parts and joints can still form disconnected linkages, immobile assemblies, or motions that violate the intended structure. We present , a kinematic-graph-centered agent harness for generating executable articulated-object programs from visual and textual specifications. The harness organizes construction, numerical execution, and revision around a shared representation of rigid parts, connection interfaces, joints, and coupling constraints. Explicit constraints beyond the joint tree allow a solver to assemble the mechanism and recover coupled motion, including nonlinear dependencies between driver and follower joints. Structural and motion evaluation then grounds model-driven revisions of the graph, interfaces, and geometry in the resulting execution evidence. By connecting structural decisions to their consequences over motion, MARC makes mechanism generation an iterative process of proposing, testing, and revising executable constraints.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Articulated Object Generation, Kinematic Graphs, Programmatic 3D Modeling, LLM Agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_S7RDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

MARC is introduced as an agent framework for generating articulated-object programs from images and text. Building upon Articraft, it incorporates socket-bound connections and a kinematic graph to represent closed loops and coupled motion. Numerical diagnostics inform a multi-stage correction workflow. Experiments on 20 objects demonstrate improvements over existing generators, with further backbone comparisons and ablations of reflection and graph-based support.

Critical strengths

1: The paper addresses a significant problem. Individually plausible parts and joints do not always constitute a functional mechanism, especially in the context of closed chains. Assessing the validity of connections throughout motion facilitates the generation of articulated objects.

2: Connecting geometry, attachment frames, and kinematic constraints through a shared representation is a reasonable design. Diagnostics tied to specific joints and sockets could make corrections more targeted than relying only on rendered images or generic execution errors.

3: The comparison with the initial Solver candidate provides some evidence that correction helps. With Kimi-K3, reflection improves valid motion coverage from 54.1% to 67.8% and overall success from 40% to 60%, while structural correctness remains unchanged.

Critical weaknesses
  1. The accompanying figures in the article are highly likely AI-generated and have not been manually refined to ensure accuracy. The AI ​​usage statement acknowledges assistance in preparing and modifying the figures, and Figure 1 explicitly shows the "mechanism generated using MARC." Which panels are direct procedural renders, and which are illustrations or generatively modified images? The authors should clearly label these and provide unmodified renders and motion sequences corresponding to the evaluation procedure.

  2. The requirements for the final evaluation lack clear definition. While Table 8 provides solver tolerances, Appendix C does not specify numerical acceptance criteria for collisions, branch continuity, or follower phase and travel. The definitions of SC and CMC depend on meeting annotation requirements, yet the adjudication process for semantic cases and the resolution of disagreements remain unexplained. These details are essential, particularly given that MARC achieves 100% SC and 98.2% CMC.

  3. The analysis of baseline results requires greater detail. Table 1 reports a 0% overall success rate for all six baseline models, with five models achieving 0% effective motion coverage. It is important to specify how many failures result from unsupported closed-loop representations, incorrect connection generation, execution adapters, or geometric interference. An object-by-object analysis of failure causes and the establishment of a baseline model with comparable solution-support mechanisms would clarify the comparative results.

Overall, focusing on closed-loop mechanisms is a valuable direction, and shared build and diagnostic interfaces offer potential practical benefits. However, the current experiments do not clearly distinguish the contributions of graph-guided corrections from those of additional mechanism support and inference-time computation. Clearer evaluation criteria, controlled comparisons, and traceable visualizations are necessary to substantiate the reported improvements.

Questions for authors

Please see the weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer S7RD

  1. The existing ablations identify the contribution of graph-guided repair.

Table 4 includes a configuration that retains the kinematic graph and numerical execution while withholding graph-based connection localization and constraint associations during repair. Full MARC achieves 67.8% VMC and 60.0% OSR, compared with 56.3% VMC and 45.0% OSR without graph-guided repair. These gains of 11.5 and 15.0 percentage points establish the contribution of graph-associated repair feedback with mechanism support retained.

The reflection ablation starts from the same initial Solver candidates. Reflective correction increases VMC from 54.1% to 67.8% and OSR from 40.0% to 60.0%, while SC and CMC remain unchanged. This directly demonstrates improvement after numerical solving, particularly in completing valid motion.

The completed PartNet-Mobility evaluation provides complementary evidence on standard articulated objects. MARC and original Articraft both export valid geometry for 77/77 inputs, while graph accuracy increases from 49.35% to 79.81% and axis error decreases from 0.13881 to 0.06377 rad. The improvement extends to inputs for which the baseline already produces usable geometry.

  1. Structural correctness, coupling correctness, and complete motion are evaluated separately.

The final evaluation is conducted independently of MARC's internal Judge. Appendix C defines the distinct acceptance criteria:

  • SC evaluates the annotated structural requirements.
  • CMC evaluates follower participation, direction, phase, and travel.
  • VMC measures the reachable valid prefix of the prescribed driver path, including required connections, limits, continuity, and prohibited interference.
  • OSR requires all critical requirements and complete valid traversal of the sampled path.

Table 8 specifies attachment tolerances of 0.0001 m and 0.001 rad. Requirement-level semantic and contact adjudications were recorded and applied before final aggregation; the completed record contains 35 decisions across 15 method-object pairs.

The reported values preserve the distinction between these criteria. MARC's 100% SC and 98.2% CMC coexist with 67.8% VMC and 60.0% OSR, with closure and interference failures retained. Table 1 also credits baseline capabilities separately: original Articraft achieves 100% ER and 24.6% CMC, while complete-mechanism success additionally requires the critical connections and valid prescribed traversal.

The standard benchmark further reports substantial nonzero baseline performance under both explicit denominator conventions. In the manuscript's manual-reference comparison, Table 2 also shows improvements across all six aggregate geometry measures, covering static geometry and approximately matched motion states.

  1. The figures have distinct explanatory and evidentiary roles.

Figure 1 provides an illustrative workshop overview. Figures 2–3 explain the representation, attachment frames, and motion constraints. Figure 4 presents native-control before/after states and explicitly retains stationary or unstable responses, including the MARC excavator's collision.

The manuscript discloses AI assistance in figure preparation. The quantitative conclusions are based on executable outputs and independent evaluation. The completed 77-object comparison provides additional quantitative evidence for articulation structure and joint parameters.

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

This paper presents MARC, an agentic framework for articulated 3D object generation that integrates a unified, executable kinematic graph with graph-guided diagnosis and reflective correction. By explicitly modeling closed-chain connections and coupled-motion constraints, MARC enables mechanism-level motion validation and targeted repair, achieving substantial improvements over existing baselines, with ablation studies supporting the effectiveness of its key components.

Critical strengths
  1. A systematic approach to complex mechanism generation. The paper integrates executable kinematic constraints, numerical motion solving, and graph-guided reflective correction into a unified framework, offering a meaningful solution to the challenging problem of generating articulated objects with closed-chain and coupled-motion constraints.
  2. Strong empirical improvements with supporting ablations. MARC substantially outperforms the evaluated baselines in valid motion coverage and overall success rate, while ablation studies provide evidence for the benefits of reflective correction, graph-based mechanism support, and graph-guided repair.
Critical weaknesses
  1. Limited evaluation scale and complexity coverage. The experiments cover only 20 articulated objects, limiting confidence in the generalizability of the reported improvements to larger and more diverse collections of mechanisms, particularly those with more complex topologies and stronger coupling constraints.
Questions for authors
  1. How does MARC scale to more complex mechanisms? Could the authors provide additional results or analysis on mechanisms with more links, closed loops, and coupled constraints, particularly regarding motion validity, correction success, and computational cost?
1 comment
AuthorsOpenReview ↗

Response to Reviewer BBzD

The completed PartNet-Mobility evaluation broadens the assessment to 77 objects from seven categories, with trained and agentic competitors evaluated by a common external scoring pipeline. It complements the original closed-chain and coupled-mechanism evaluation with standard articulation-structure and joint-parameter measurements.

The mechanism evaluation directly measures correction effectiveness. Starting from the same initial Solver candidates, reflection increases VMC from 54.1% to 67.8% and OSR from 40.0% to 60.0%. With the graph and numerical execution retained, removing graph-guided repair reduces VMC to 56.3% and OSR to 45.0%. These comparisons support both the effectiveness of reflective correction and the contribution of graph-associated repair feedback.

On PartNet-Mobility, MARC achieves 79.81% graph accuracy, 68.22% full-set Range-IoU, and 76.85% matched-joint Range-IoU. Together, the two evaluations support the shared graph's utility for articulated-object construction and targeted mechanism correction.

Reviewer_RuGiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a harness that allows SOTA LLMs to generate articulated objects conditioning on the input image + text in zero-shot fashion. The main idea is to let LLMs operate in the kinematic graph domain, and design a self-refinement loop that allows the agent to gradually design and assemble object parts and refine them based on the automated judge process.

Critical strengths

The method is claimed to work zero-shot on small scale evaluations. The same harness should work with different SOTA LLMs and improve over time with the capabilties of the LLMs. The authors designed various metrics to measure different performance aspects.

Critical weaknesses

The experiments are operated in non-standard benchmarks with a tiny scale, which makes it hard assess how well it performs in practice and how much better it is comparing to training based methods.

  • Only 20 objects are used in the eval, while there are benchmarks with thousands of objects in existing articulated object generation literatures. I think the analysis should extend at least to much larger than the current level and in the same standard benchmarks as existing training-based methods, so that the proposed method can compare against both agentic and non-agentic approaches.

Also the cost of this approach for generating an average object needs to be evaluated as well, both at speed and token cost level. Since the training-based approaches can be much faster potentially due to the one-shot generation.

Questions for authors

Is there any edge case or pattern that the existing/proposed method will fail? It would be interesting to know the boundary of when this type of methods will work in practice.

1 comment
AuthorsOpenReview ↗

Response to Reviewer RuGi

The completed evaluation now includes a standard 77-object, seven-category PartNet-Mobility benchmark, comparing MARC with the trained ArtLLM model and the agentic methods Articulate-Anything and original Articraft. The common response provides the configurations, results, and both statistical conventions.

MARC exports valid geometry for 77/77 inputs and achieves the highest full-set joint type accuracy (80.69%), Range-IoU (68.22%), and graph accuracy (79.81%), together with the lowest axis error (0.06377 rad). Its matched-joint Range-IoU also remains highest at 76.85%. These results extend the evidence from the mechanism-focused evaluation to a standard articulated-object benchmark with trained and agentic competitors.

The manuscript also identifies concrete failure modes. The folding desk encounters interference before completing its requested travel, and the excavator fails its bucket-closure check. These failures remain reflected in VMC and OSR even when structural and follower-response requirements are satisfied. The evaluation distinguishes construction correctness, local coupling behavior, and complete valid motion.

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

Related papers

Discussion (0)

Sign in to comment.