acceptodds
Under review as a conference paper at ICLR 2027

AdaCraft: Evolving Reliable Skills for Articulated 3D Generation

Abstract

Generating a 3D articulated asset requires geometry that remains coherent as its parts move. Agents can repair individual construction programs, yet errors in joint frames, clearance and attachment recur across objects. Reusable skills could retain these corrections, but authoring and validating them requires expertise and repeated execution. This motivates automatically learning such guidance directly from execution feedback accumulated across generation attempts. However, automatically evolving skills introduces the challenge of stability and update adaptivity. Useful repairs must remain stable across later revisions, and the scope of each update must adapt to its effects on different mechanisms. We propose **AdaCraft**, a framework that addresses these challenges through Motion Memory and an Adaptive Edit Budget. Motion Memory retains geometric conditions and evaluated repairs, and uses a Homological Motion Signature to retain connection margins and bottleneck witnesses for required poses on a sampled motion graph. The Adaptive Edit Budget uses variation in functional improvements, execution uncertainty, geometric dependencies and motion fragility to control how many rules may change. Across 240 test tasks, AdaCraft achieves 75.1% functional yield compared with 58.1% for the strongest control, while reducing deployment calls per functional asset from 28.9 to 16.8.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D asset generation, articulated objects, skill evolution, adaptive optimization

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

The paper tasks coding agents to construct articulated assets, and aims for them to improve after each asset by accumulating information as prompts/skills to help with future assets. This forms a type of skill-learning, where a checkpoint is a set of skills that has been developed by working on a set of training tasks, during which the skills change but the LLM weights are unchanged. AdaCraft is a process for updating skills after each task that preserves correctness of articulated assets and limits the rate of change. The authors show that the final AdaCraft-optimized skill set produces more frequently correct articulated assets than either the base agent or an agent with unstructured self-adaptation of its skills.

Critical strengths

Testing the ability of agents to look at past errors and develop skills that help it in the future is interesting, and the work shows that agents' ability to do this can be significantly improved, over and above their default behavior when doing this.

Critical weaknesses

ArtiCraft is only included in qualitative results not quantitative, despite that it is to my knowledge the strongest harness for articulated object generation, and includes similar functionality checks as are used to develop AdaCraft. I could interpret the paper as a more nuanced contribution studying specifically the ability to make prompts from scratch, but at the moment its components seem very similar to an articulated object harness like articraft, just retained across sessions.

Questions for authors

Would you be able to summarize what the agent discovered to be useful? What does the final AdaCraft configuration contain after gathering across many tasks? How do they compare to the manually designed prompts of ArtiCraft?

Can AdaCraft be compared quantitatively to ArtiCraft? I would hope this can be done using its paper results or its asset corpus without running its harness. If AdaCraft is not expected to be better than ArtiCraft, then is the

Where does ArtiCraft fall on the Fig5a pareto front? Would it be a point with high functional yield and 0 optimization tokens?

NOTE - Questions from this line and onwards are non-essential:

How many tokens total were spent accumulating the skills used in the final AdaCraft system? Is it possible that if ArtiCraft were given the same additional tokens at train time it might outperform AdaCraft, which already spent its extra budget on other tasks? At what number of test examples would the upfront investment in AdaCraft pay off? Note: this question is not essential.

Would it be reasonable to have a baseline similar to Naive accumulation, but where agents are just told to test if their skill is good by writing code? Testing the skill via code is a generalization of the current harness. Or is this already present in naive accumulation and just doesnt occur successfully?

3 comments

We thank the reviewer for recognizing the value of learning reusable skills from past errors and AdaCraft’s improvements over agents’ default skill-accumulation behavior. Below, we clarify what the system learns, its relationship to ArtiCraft, and the corresponding performance and computational costs.

Q1. What does the final AdaCraft configuration contain and how useful they are

The initial task-skill library has four modules: geometry, joint structure and frames, motion clearance, and physical attributes/export. These supplement the underlying ArtiCraft harness without changing its frozen generation model or fixed tools.

The initial rules cover basic operations such as part naming, selecting joint types from required motions, choosing hinge pivots, setting gaps, inspecting motion endpoints, and assigning inertia.

AdaCraft subsequently revises these instructions based on cross-object evidence. The important distinction is that the revised library contains persistent applicability-conditioned procedures, whereas ArtiCraft's original harness performs construction and verification within an individual episode.

Table R10 summarizes representative changes from the initial task-skill library to the final selected library. The accepted seed-42, round-3 J2/J3 revision in Figure 7 and Table R4 provides a concrete example: it restricts hinge placement to appropriate mechanisms, aligns the hinge axis with the stationary seam, and specifies limits that preserve the required travel. Its compatible-objective gain is ΔQ = 0.021, and the revised rules entered the active library after acceptance.

Table R10. Evolution of the construction-skill library.

ModuleInitial procedureFinal learned procedure / change
GeometryName every part and joint by its required role.Construct and name bodies according to their functional roles. Keep the stationary housing and independently moving components as separate rigid bodies; attach decorative geometry to its owning body. For example, a cabinet has housing and door_0 connected by door_0_joint, while its recessed pull belongs to door_0. Derive the door dimensions from the opening dimensions and specified clearance, and check both the housing bounds and the assembled object’s reference bounds.
Joint structureUse prismatic joints for sliding and revolute joints for swinging or spinning; place revolute pivots on a hinge edge.Restrict hinge placement to edge-hinged swing/fold mechanisms. Place the axis on the stationary seam shared with the parent: the side edge for a side-opening door, bottom edge for a drop-down door, back edge for a lid, or connecting pin for a folding link. Exclude centered knobs and rotors from this hinge rule. Specify finite, task-dependent limits: sliders must cover the requested translation, and hinges must cover the required opening angle. Verify the motion direction and both endpoints rather than assigning one generic angular range to all revolute joints.
Motion clearanceLeave a gap of 0.005L0.005L; check forbidden contacts at rest and at motion endpoints.Check intermediate configurations as well as endpoints. Begin with the specified construction gap, then evaluate the moving part against the housing and neighboring components at minimum, midpoint, and maximum travel. For a hinged door, offset its closed pose from the parent face and check whether the panel sweeps into the frame as it opens. Use the diagnostic’s offending part pair and joint configuration to guide changes to the pivot or geometric offset. Recheck the required travel after each change; shortening motion below the task requirement does not resolve the failure.
Physical/exportCompute mass from volume and density, and approximate inertia from the part’s bounding box.Give every rigid body physical properties and verify a successful export. Assign material density or explicit mass properties; compute m=ρVm=\rho V and bounding-box inertia using consistent units—for example, Ixx=m(dy2+dz2)/12I_{xx}=m(d_y^2+d_z^2)/12. Compile early, use compile diagnostics to repair the program, and check that the exported manifest contains the required bodies, joints, articulation, and reference state. Evaluate the exported asset: a generated program or preview alone is insufficient, and an episode without a successful export counts as a failure.

We have added a summary of the learned procedures in Section 4, ‘What the learned library contains,’ covering geometry, joint structure, motion clearance, and physical attributes/export.

Q2. How AdaCraft differs from ArtiCraft:

The “Fixed agent” baseline in Table 1 evaluates ArtiCraft under our shared model, benchmark, and resource settings, without our additional task-skill library. Here, “fixed” means that no persistent task-skill library is updated across objects; ArtiCraft’s testing and repair within each generation episode remain active. We agree that the original label obscures the ArtiCraft comparison and have renamed it to “ArtiCraft harness (fixed agent)” in Table 1 in the revised paper.

Our contribution is reliable cross-episode skill evolution, enabled by three complementary components: (1) Motion Memory, which retains successful repairs and counterexamples with explicit mechanical applicability conditions; (2) Homological Motion Signatures, which capture motion connectivity and identify clearance bottlenecks beyond individual pose checks; and (3) Adaptive Edit Budgets, which regulate revision scope based on transfer variability, execution uncertainty, geometric dependencies, and motion clearance. Together, these components enable reusable skill improvement while preserving previously acquired capabilities. Overall, the idea is to give an "Adam optimizter" for the coding agent.

Table R11. Quantitative comparison under the same evaluation protocol.

MethodFunctional yield ↑Calls / functional asset ↓Optimization tokens
ArtiCraft harness (fixed agent, no task skills)36.2%48.10
ArtiCraft + fixed initial task skills40.7%42.50
SkillAdam + shared gate58.1%28.9214.7M
ArtiCraft + AdaCraft skill evolution75.1%16.885.4M

AdaCraft improves functional yield by 38.9 percentage points relative to the fixed-agent condition and reduces deployment calls per functional asset from 48.1 to 16.8.

More importantly, against SkillAdam with the same acceptance gate, AdaCraft improves yield by 17.0 percentage points, demonstrating benefits beyond the shared generation harness or candidate-validation mechanism.

Q3 Where does ArtiCraft fall on the Figure 5(a) optimization–performance curve?

The fixed ArtiCraft condition requires zero cross-episode optimization tokens. It would sit on the Y axis. However, it does incur its normal generation and repair costs during deployment.

Q4 Is the additional optimization cost justified, and could ArtiCraft benefit from equivalent compute?

We agree that training-time cost and deployment-time efficiency should be distinguished.

AdaCraft uses 85.4M optimization tokens in the reported experiments, including diagnosis, candidate generation, validation, and rejected revisions. This is upfront expenditure. Table R12 summarizes optimization and deployment cost:

Table R12. Optimization and deployment cost.

MeasureArtiCraft fixed agentSkillAdam + gateAdaCraft
Optimization tokens0214.7M85.4M
Functional yield (%)36.258.175.1
Deployment tokens / functional asset1.443M0.867M0.504M
Successful objects needed to recover optimization cost (vs. ArtiCraft)—37391

ArtiCraft could also benefit from additional generation or repair compute. However, SkillAdam + shared gate already optimizes skills using the same harness and consumes 214.7M optimization tokens, yet achieves 58.1% functional yield, compared with 75.1% for AdaCraft. This supports the efficiency of our skill-evolution strategy.

Q5 Code-based testing?

Yes, code-based testing is definitely possible. And, in fact, they are already available through the underlying ArtiCraft harness! However, these local tests frequently fail to produce functionally valid articulated assets and do not establish whether a skill learned from one object will generalize to others.

As shown in Table 1, Naive Accumulation achieves only 39.4% functional yield, meaning that 60.6% of generation attempts fail the prescribed structural or motion checks. Simply accumulating corrections, even with access to execution feedback, is therefore insufficient.

Table R13. Functional yield and failure rate under code-based testing.

MethodFunctional Yield ↑Failure Rate ↓
Fixed agent (ArtiCraft harness)36.2%63.8%
Naive Accumulation39.4%60.6%
SkillAdam + shared gate58.1%41.9%
AdaCraft75.1%24.9%

Importantly, testing an individual generated asset is different from validating a reusable skill across objects. A correction that passes local code-based tests may still fail under different joint configurations or invalidate previously successful mechanisms.

AdaCraft addresses this limitation through mechanically conditioned Motion Memory, adaptive revision budgets, and cross-object validation.

Dear Reviewer bXqy,

Thank you again for your thoughtful questions and constructive feedback. Your input has helped us significantly strengthen our work, clarify the relationship to ArtiCraft, and explain what AdaCraft learns across tasks. As the discussion period is nearing its end, we wanted to follow up and invite your thoughts on these clarifications.

Thank you for your time and consideration!

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

AdaCraft evolves reusable construction skills for articulated 3D asset generation while keeping the generation model fixed. Motion Memory stores repair evidence and mechanical applicability conditions, while Homological Motion Signatures summarize connectivity and bottleneck configurations on sampled motion graphs. An adaptive edit budget limits revision scope, and candidate updates are validated on compatible, boundary, and protected tasks. Across 240 held-out tasks covering new instances, categories, and mechanism compositions, the paper reports 75.1% functional yield versus 58.1% for SkillAdam with the same acceptance gate.

Critical strengths
  1. The framework combines reusable repair evidence, explicit applicability conditions, motion diagnostics, and constrained revisions to address recurring construction errors and regression across tasks. Recording when a repair applies and checking both transfer and retention makes the approach relevant to articulated generation, where a correction for one mechanism can disrupt another.

  2. AdaCraft improves overall functional yield from 58.1% to 75.1% over SkillAdam with the same acceptance gate (Table 1), supporting benefits beyond candidate validation alone. Component ablations support the usefulness of structured memory and adaptive editing (Table 2). Table 3 also provides an informative distinction: motion signatures outperform minimum clearance on tasks permitting alternative routes, while matching it on prescribed chains, where the connection margin reduces to minimum path clearance.

Critical weaknesses
  1. Insufficient operational detail on the connection between motion evidence and rule revision. Section 3 and Appendix A.3 specify the memory schema, retrieval criteria, storage limits, and general update policy. However, it remains unclear how asset-level motion signatures are associated with particular rule IDs and incorporated into the optimizer input to produce concrete rule edits. Figure 2 provides a schematic illustration rather than a worked revision trace. This gap limits reproducibility of the central mechanism connecting motion diagnostics to reusable skill updates.

  2. Incomplete disclosure of benchmark coverage and retention evidence. Appendix A.1 provides task counts and split policies, but not concrete lists of held-out categories and joint arrangements or representative task manifests specifying the required motions. This limits assessment of the scope and difficulty of the generalization setting. The appendix also states that eligible retention-audit subset sizes and mechanism coverage are reported, but I could not locate those results; these denominators are needed to interpret the breadth of the reported 98.1% retention. Likewise, the paired 95% confidence intervals for method differences described in Appendix A.6 are not shown. These omissions limit assessment of coverage and uncertainty, although Table 1 does report a substantial yield improvement with means and standard deviations across three runs.

Questions for authors
  1. Could you clarify how asset-level motion signatures are associated with specific rule IDs by providing one complete revision trace, including the diagnostic evidence, retrieved memory records, optimizer input, before/after rule text, and acceptance outcome? This would help assess whether the central skill-update procedure can be reproduced.

  2. Could you provide the held-out category and joint-arrangement lists, together with representative task manifests specifying required parts, joint roles, and motions? This would clarify the coverage and difficulty of the generalization benchmark.

  3. What are the eligible retention-audit subset sizes and mechanism coverage in each run? Could you also provide the paired 95% confidence interval for the overall functional-yield difference between AdaCraft and SkillAdam with the shared gate, as described in Appendix A.6? These results would clarify the coverage of the retention claim and the uncertainty of the primary comparison.

5 comments

We thank the reviewer for recognizing the benefits of mechanically conditioned memory and constrained revisions, the shared-gate gains, and the supporting ablations.

W1/Q1. Connecting motion evidence to rule revision

AdaCraft associates evidence with rules through rule activation and mechanical relevance, rather than task category alone:

  1. Diagnostics. Each generation records applied rule IDs and part roles. Evaluation returns failed checks, implicated part/joint IDs, tested configurations, and motion records.
  2. Association. Memory matches diagnostics to the activated part/joint and affected checks; joint IDs link motion summaries to rules. Records retain supporting cases, counterexamples, motion summaries, and revision history.
  3. Retrieval. Joint type, graph motif, part role, and frame relation select relevant records. The optimizer receives the current module, diagnostics, applied rule IDs, motion summaries, retrieved evidence, and constraints to preserve.
  4. Bounded edit. The optimizer declares target rule IDs. Transfer variability, execution uncertainty, dependency reach, and clearance risk determine the edit budget; a second call produces the edits within it.
  5. Acceptance. Fresh generations on compatible, boundary, and protected cohorts determine whether the candidate replaces the active rules.

Table R4. Accepted seed-42, round-3 revision.

ComponentRecorded informationRound-3 trace
Applied skillRule ID and part roleSeed 42, round 3; J2 controls hinge applicability and placement, and J3 specifies required joint limits. Part/joint identifiers: door_0 / door_0_joint, with housing as the stationary parent.
FailureFailed check and jointDiagnosis identified insufficient joint limits and contacts between moving parts and their housings. Diagnostic: door_0–housing clearance fails at q = 0.785 rad, and door_0_joint stops short of the required opening.
HMSConnection margin and bottleneckHMS record: T_J2 stores μ = −0.001 on the prescribed door-motion chain and bottleneck (door_0, housing, door_0_joint, q = 0.785 rad). The joint ID links the diagnostic to the activated rule.
Motion MemorySupporting and negative casesMemory records: P_J2 = mem-J2-hinge-01, a stationary-seam repair; N_J2 = mem-J2-rotor-02, a centered-rotor scope counterexample; T_J2 = the bottleneck record above. After validation, H_J2 records the accepted round-3 revision.
Retrieved evidenceMatched records and constraintsRetrieved IDs: mem-J2-hinge-01, mem-J2-rotor-02, mem-J3-travel-01. Match compatible hinge role and frame relation; retain the rotor case as a scope counterexample and preserve constraints from successful constructions.
Optimizer inputProposed optimizer input textOptimizer input: “Targets: J2, J3. Use the joint-limit and moving-part/housing contact diagnostics, compatible hinge repairs, and scope counterexamples. Restrict hinge placement to edge-hinged swing/fold mechanisms; exclude centered rotors and sliders from this rule. Set limits from required travel and verify endpoints and intermediate poses. Edit only J2/J3 within the rule and changed-token caps; return revised procedures and applicability conditions.”
RevisionOriginal → revised procedureJ2: “All revolute joints; place the pivot on the hinge edge at the outer face the part opens toward” → restrict to edge-hinged swing/fold mechanisms, exclude centered rotors and sliders, and place the axis on the stationary hinge seam with closed/intermediate/open-pose clearance checks. J3: no dedicated rule → finite task-dependent limits with required-travel and clearance checks.
Adaptive budgetTransfer, uncertainty, dependency, clearanceTargets: J2 and J3. The protocol retains the initial two-rule cap until the module has three completed comparisons spanning eight distinct tasks; the separate changed-token cap is 1,024. The round-3 proposal changes two rules.
ValidationCompatible/boundary/protected resultsEight compatible objects × two generations: functional validity 7/16 → 9/16; compatible-objective gain ΔQ = 0.021 > 0.02. Gate outcome: accepted. The validation protocol also evaluates boundary and protected cohorts.
DecisionAccepted or retainedAccept the round-3 candidate and update the active J2/J3 rules. The accepted procedures can be retrieved for later objects satisfying their mechanical applicability conditions.

Table R4 and Figure 7 describe the same accepted revision. Our response to Reviewer 7uqz describes its later reuse on a mailbox. Section 4 now explains the diagnostic-to-rule association and validation decision; Figure 7 shows the original and revised rules. Part 2 provides the supporting ablations, benchmark coverage, and retention results.

W1/Q1 (continued). Supporting ablations

The motion-connectivity gains are supported by paper Tables 2–3:

Table R5. Contribution of motion-aware optimization (paper Tables 2–3).

ComparisonBaselineAdaCraft / HMSImprovement
Functional yield without vs. with HMS68.2%75.1%+6.9 pp
Functional yield without vs. with clearance penalty71.0%75.1%+4.1 pp
Alternative-route yield: minimum clearance vs. HMS58.6%72.7%+14.1 pp
Prescribed-chain yield: minimum clearance vs. HMS76.9%76.9%0.0 pp
Dense-check failure rate: minimum clearance vs. HMS8.8%5.7%−3.1 pp

W2/Q2 benchmark coverage and retention evidence:

We agree that these lists should be provided explicitly.

Table R6. New-category generalization.

Mechanism groupHeld-out categories
Revolutesafe, mailbox, chest, lounge chair, turnstile, faucet, air bike, rolling walker
Prismaticfile cabinet, dresser, sliding door, car sunroof, utility knife
Mixedtelescoping boom, desk, headphones, radar dish, folding scooter, louvered shutter

The new-composition partition instead uses familiar categories with held-out arrangements:

Table R7. New-composition generalization.

CategoryHeld-out arrangement(s)
Cabinetdouble, stacked
Microwavedoor-right
Ovenstrip-bottom
Storage boxside-hinge
Knob panelgrid
Drawer cabinetsplit
Nightstanddrawer-bottom
Sliding windowslide-left
Refrigeratordoors-bottom
Folding armfold-right
XY stagey-below

The task manifest specifies the required parts, parent-child joints, joint types and axes, motion ranges, required trajectories, and contact exemptions. The protocol also fixes the reference length and motion graph independently of generated assets.

Table R8 gives representative manifest entries aligned with the category and arrangement lists above; nc_mailbox_001 is also the later-use example in our response to Reviewer 7uqz.

Table R8. Representative manifest records.

SplitTask IDRequired parts/jointsRequired motionContact exemptions
New instancesni_drawer_cabinet_001Drawer cabinet: housing + drawer_0; drawer_0_joint is prismatic along the rail axis (2 parts, 1 joint).Drawer translates 0–0.20 m along its rail axisDesignated drawer–rail guide interfaces only
New categoriesnc_mailbox_001Mailbox: housing + door_0; door_0_joint is revolute about the vertical side-hinge axis (2 parts, 1 joint).Door rotates 0–1.57 rad from closed to open; check the prescribed single-joint motion chain.Designated door-hinge bearing interface only.
New compositionsncomp_cabinet_double_001Cabinet, double-door arrangement: housing + door_left + door_right; two revolute side-hinge joints with axes oriented for outward opening (3 parts, 2 joints).Each door rotates 0–1.57 rad; reach the both-open target from both-closed on the fixed two-joint motion graph.Designated left/right door-hinge interfaces only; no door–door contact exemption.

Q3. Retention-audit details and confidence intervals

We agree that the reported 98.1% retention requires explicit denominators. The retention-audit pool contains 24 tasks (8 revolute, 8 prismatic, and 8 mixed), but eligibility is defined separately within each initial run by requiring both initial generations to be functional.

The selected subset is fixed throughout optimization. The optimizer does not receive audit trajectories or retention scores.

The table below reports retention coverage and final retention by seed. Each eligible task receives two fresh generations. Aggregation uses unrounded per-seed rates with equal run weights.

Table R9. Retention coverage and uncertainty.

QuantitySeed 42Seed 43Seed 44Mean ± SD
Eligible audit tasks24/2418/2418/24—
Revolute / prismatic / mixed8 / 8 / 86 / 6 / 66 / 6 / 6—
AdaCraft retention (%)100.0 (48/48)94.4 (34/36)100.0 (36/36)98.1 ± 3.2
SkillAdam + gate retention (%)97.9 (47/48)91.7 (33/36)91.7 (33/36)93.8 ± 3.6

For the primary functional-yield comparison, the reported means are 75.1% for AdaCraft and 58.1% for SkillAdam + shared gate, a 17.0 percentage-point improvement.

We computed a paired 95% bootstrap confidence interval by resampling runs and source tasks within each test split, using the same resampled indices for both methods and keeping repeated generations grouped by task. AdaCraft improves overall functional yield over SkillAdam by 17.0 percentage points, with a 95% confidence interval of [+11.8, +22.3] percentage points.

Dear Reviewer NMjn,

Thank you again for your thoughtful feedback, which helped us strengthen the paper’s reproducibility and evaluation. We would welcome your thoughts on the added revision trace, benchmark details, retention analysis, and the confidence intervals for the improvement over SkillAdam. As the discussion period approaches its end, we would be grateful if you could let us know whether these additions address your concerns.

Thank you once again for your time and thoughtful review.

Reviewer_7uqzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AdaCraft improves an articulated 3D generation agent by updating a reusable skill library while keeping the language model fixed. It stores repair experience with conditions describing when each rule applies, summarizes motion connectivity and clearance, and controls how many rules can change at each update. Candidate rules are tested on applicable, boundary, and previously successful tasks before being accepted. Experiments on 240 test tasks report higher functional yield and fewer deployment model calls than existing skill optimization methods.

Critical strengths

The paper addresses a useful problem: a repair that helps one object can fail on another, and later rule changes can undo earlier improvements. Recording mechanical conditions and testing new rules through fresh generations are reasonable ways to reduce these failures.

Critical weaknesses
  1. The benchmark needs to be clearer. The paper gives task counts and split rules, but does not list the source assets or category splits. The retention result also uses only initially successful tasks from a pool of 24, without reporting the actual subset size for each run. These details are important for judging the generalization and retention claims.

  2. The paper does not show a concrete example of a skill before and after revision, or how the revised skill helps generate another object. Since reusable skill evolution is the main contribution, this makes it difficult to understand what the agent actually learns.

Questions for authors
  1. What do J1 and C2 exactly mean in Figure 1?

  2. Could you show one actual skill revision, including the original rule, diagnostic feedback, revised rule, validation outcome, and its later use on a different object?

6 comments

We thank the reviewer for recognizing the importance of mechanically grounded skill evolution and fresh-generation validation. Below, we provide additional benchmark and retention details, present a concrete skill-revision example, and clarify the notation in Figure 1.

W1 benchmark clarification and retention coverage:

We agree that the task coverage should be directly reported. The evaluation uses 480 tasks, including 240 held-out test tasks. The primary benchmark is constructed from parameterized articulated-object templates, with fixed structural and motion requirements. We report the benchmark partitions below.

Table R1. Benchmark partitions.

PartitionTasksRevolutePrismaticMixed
Diagnosis120404040
Promotion36121212
Retention audit24888
Development60202020
New instances80303020
New categories80303020
New compositions80303020

For capability retention, we select tasks for which both initial generation attempts succeed and fix this eligible subset for each run. The same initially established tasks are evaluated through fresh generations at later checkpoints. Each eligible task receives two fresh generations; aggregate retention is the equally weighted run mean ± sample SD, calculated from the exact success fractions.

Table R2. Initially established retention tasks.

Run seedEligible / 24RevolutePrismaticMixedAdaCraft retention
4224 / 24888100.0% (48/48)
4318 / 2466694.4% (34/36)
4418 / 24666100.0% (36/36)
Mean ± SD————98.1 ± 3.2%

W2 / Q2 An example of a skill before and after revision:

Seed 42, round 3 provides a concrete accepted revision of J2/J3. Diagnosis identified insufficient joint limits and contacts between moving parts and their housings. The table summarizes the original and accepted instructions, using the objects in Figure 7 to explain how the revised procedures apply.

Table R3. Accepted J2/J3 instructions before and after revision.

Rule / purposeBefore revisionAccepted revision (summarized)
J2: Mechanism scopeAll revolute joints; the same hinge-edge placement instruction can apply to swinging doors and centered spinning parts.Use the edge-hinge procedure for swinging/folding doors, lids, flaps, and links. A side-opening cabinet door qualifies because it rotates about an edge; a ceiling-fan rotor does not, because it spins about its center. Exclude centered spinning controls and prismatic sliders from this rule.
J2: Hinge construction“Place the pivot on the hinge edge at the outer face the part opens toward.”Place the joint axis on the hinge seam shared with the stationary parent. For a side-opening cabinet door, use the vertical side seam; for a laptop lid, use the rear hinge seam. Provide clearance in the closed pose, then check the moving part against its housing at closed, intermediate, and required open poses.
J3: Joint limitsNo dedicated J3 rule for joint limits.Read the required motion from the task and set explicit finite limits that cover it. For a storage-box lid, the upper limit must reach the specified opening angle. Verify travel and clearance at both endpoints and intermediate poses; shortening the required opening to avoid contact does not satisfy the task.

Validation and acceptance. On eight compatible objects with two generations per object, the complete J2/J3 revision improved functional validity from 7/16 to 9/16 and increased the compatible objective by ΔQ = 0.021, exceeding the 0.02 gain threshold. The gate accepted the revision and updated the active library. The table therefore describes accepted construction procedures.

Later use on another object. The frozen library reuses the accepted hinge procedure on held-out mailbox task nc_mailbox_001. J2 locates the mailbox’s stationary hinge seam and checks the door against its housing along the opening motion; J3 preserves the required 0–1.57 rad travel. Matched generations with the old and revised libraries give 0/2 → 2/2 functional outputs. This shows the procedure being reused on a different object with compatible hinge mechanics. Mailbox feedback does not update the library.

We have added the accepted round-3 revision in Section 4 and Figure 7, showing the original and revised J2/J3 rules, the compatible-cohort improvement, and the acceptance decision.

Q1 What do J1 and C2 represent in Figure 1:

Figure 1 uses J1 to represent a joint-related construction constraint that should be preserved, and C2 to represent a clearance-checking rule that can be refined without modifying that constraint.

Dear Reviewer 7uqz,

Thank you for your thoughtful review and for recognizing the value of mechanically grounded skill evolution in AdaCraft. Your feedback helped us strengthen the paper and clarify important details. We would welcome the opportunity to discuss any remaining questions or concerns. Please let us know whether our responses address the points you raised or whether further clarification would be helpful.

We sincerely appreciate your feedback and consideration.

Thank you for your reply. I have a question about Figure 6. For the microwave example, why do the back views look the same across different methods?

And for J1 and C2 in Figure 1, I think you should add the explanation in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.