acceptodds
Under review as a conference paper at ICLR 2027

Understanding Co-Design: Exploiting the Asymmetry between the Designer and Controller

Abstract

Robot co-design jointly optimizes morphology and control against the same final objective, but does that mean the two optimizers should behave alike? To answer this, we create a test environment and explore it through the lens of three new metrics: spread, exploration, and generalization. By varying these properties independently, we find that effective co-design requires opposite behavior from each optimizer: the controller requires broad generalization to new designs and suffers from wide sampling, whereas the designer barely benefits from generalization but must sample the design space broadly to find good designs. We show that the source of this asymmetry is due to the sequential structure of co-design in which the body is chosen before the controller acts, and that it can be explained through the Expected Value of Perfect Information. Our test environment also shows that broader morphology search can discover better designs but may simultaneously worsen the selection when the controller cannot evaluate unfamiliar bodies reliably. We observe optimization profiles in existing co-design methods, and combine our findings to construct SCOPER which is tailored to exploit the optimizer asymmetry by maintaining a broad morphological search space, and using representation learning to increase controller generalization. In our experiments, SCOPER demonstrates the value of following these principles by outperforming prior co-design methods task performance by  50% and halving runtime.

Primary area:
reinforcement learning
Keywords:
Morphology-Control Co-Design, Reinforcement Learning, Morphology Optimisation, Design Optimisation, Bilevel Optimisation, Robot Learning, Representation Learning, Joint Optimisation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_frE2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the asymmetric roles of the designer and controller in morphology–control co-design. The authors introduce three metrics, spread, exploration, and generalization, to analyze their optimization behaviors. Based on the findings, they propose SCOPER, which achieves improved final performance and wall-clock efficiency on a modular robot locomotion task.

Critical strengths
  1. The paper investigates an important question about the different optimization behaviors required by the designer and controller, offering a meaningful perspective beyond developing another co-design algorithm.
  2. The proposed metrics provide a useful framework for analyzing and comparing different co-design methods.
Critical weaknesses
  1. Efficiency comparison. SCOPER uses 201M environment samples, compared with 50M for BodyGen (Table 2). While SCOPER achieves better wall-clock efficiency, its advantages in sample efficiency are therefore unclear. It would also be helpful to better distinguish the algorithmic improvements from the benefits of its highly parallelized implementation.
  2. Limited task diversity. The experiments cover only one Modular Ant locomotion task. Different tasks may have very different morphology fitness landscapes, potentially changing the importance of designer spread and controller generalization.
Questions for authors
  1. The definition of ρC(k)\rho_C(k) in Eq. (5) may reflect differences in morphology potential in addition to controller generalization. For example, lower performance on an unfamiliar body could result from its lower potential rather than limited controller adaptation. Please correct me if I have misunderstood this metric.
  2. Eq. (6) uses the optimal response (a∗(d))(a^*(d)), assuming that the controller can optimally respond to every sampled design. How should this analysis be interpreted when the controller has limited adaptation ability?
  3. Could the authors provide visual comparisons of the morphologies discovered by SCOPER and baseline methods? In particular, showing representative designs at different training stages would help illustrate how their morphology search behaviors differ and whether BodyGen indeed converges prematurely.
2 comments
AuthorsOpenReview ↗

Reviewer frE2 Response

We want to thank you for your time and constructive feedback on our paper.

@1) Interpretation of ρC\rho_C

Your interpretation is correct. As detailed in Common Response C1, we now refer to ρC\rho_C as controller competence transfer and clarify that it measures realised performance retention over each method's own morphology-search distribution, rather than morphology-independent controller ability. We additionally report a fixed-bank evaluation over the same 56 morphologies, using body-specific specialists as a common reference for cross-method comparison.

@2) Eq. (6) with limited controller adaptation

You are correct that Eq. (6) assumes the optimal design-conditioned response a∗(d)a^*(d), whereas practical controllers may not realise this response. We retain Eq. (6) to describe the structural value available from conditioning on the realised design, and add Eq. (7) to account explicitly for the shortfall of the controller available during optimisation.

When controller competence is limited, this shortfall offsets part of the value available through conditioning: a high-potential design can receive low realised fitness and therefore be passed over during selection. The commitment-gap experiment demonstrates this mechanism empirically: broader designer search discovers higher-potential designs, while stronger controller competence transfer narrows the gap between what is discovered and what is ultimately selected.

We therefore distinguish the structural result of Eq. (6) from the empirical sampling-breadth preference: EVPI explains why designer-side uncertainty can be more recoverable, while the controlled experiments determine when broader designer sampling is actually beneficial.

@3) Visually comparing SCOPER to the baselines

We add representative early, middle, and final committed morphologies for SCOPER, NGE, and BodyGen in App. D, Fig. 6, directly addressing your request.

Qualitatively, NGE trends toward simpler morphologies over training, BodyGen concentrates early on a stable morphology and changes little thereafter, while SCOPER continues to modify its morphology after initially finding a stable design. We present these visualisations as qualitative context for the optimisation profiles measured in the main experiments, rather than as evidence for a particular causal explanation from a single run.


We hope these clarifications and additional results address your concerns, and we would greatly appreciate your reconsideration of the paper's rating if so.

I thank the authors for their detailed response. Most of my concerns have been addressed, and I will maintain my positive rating.

Reviewer_Gi2hOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors investigate the problem of co-designing a robot's body and control policy, and whether the designer and controller should behave differently given the bi-level nature of the problem. Three metrics (spread, exploration, and generalization) were defined and applied to both the designer and the controller. In a toy environment the authors found that the spread of the designer had positive effects on overall performance whereas spread in the controller reduced performance and that when the designer search is broad but the controller does not generalize well performance can be lowered relative to a narrow design search. The authors then apply similar, but newly defined, metrics in a robot co-design task and evaluate baseline methods NGE and BodyGen. They propose SCOPER, a transformer-based architecture with a shared representation for both designer and controller with an entropy weight for the designer specifically. They report an approximately 2x return compared to the next best method in approximately half the wall-clock time, though it requires 4x more environment samples.

Critical strengths

How should morphology and policy optimizers function to best navigate the bi-level structure of the robot design? This is an important and understudied question. This is underscored by a reasonable prior work section that highlights limits and challenges of using shared controllers. The problem setting is well-articulated and the authors correctly point out that the actual optimal fitness of a design may differ from the fitness obtained through a shared policy. The concept of the commitment gap is insightful and clear. That large design spread may only be beneficial in a regime where the controller can generalize well across a diverse set of designs is immediately intuitive. That final performance of the selected design can decrease when the controller generalization cannot keep up with design spread should be studied further, preferably in a robotics experiment.

Critical weaknesses

There are a number of inconsistencies between the authors' observations about co-design, the chosen metrics, the baseline methods, and the authors' claims.

The toy environment proposed in Section 3.3 employs metrics for measuring spread, exploration, and generalization that are unique to its setup and are not used in the robot design experiments later in the paper. Thus, the asymmetry between designer and controller identified in this section cannot be understood to automatically apply to a more complex bi-level co-design problem such as brain-body joint optimization. In the toy example, design fitness is evaluated as the expected fitness over actions sampled from the controller with standard deviation equal to the controller’s spread. So the controller spread sweep (Figure 2) measures both the spreadʼs impact on search as well as induced evaluation noise.

The generalization of the controller, ρc, mixes the controllerʼs actual generalization ability with morphology potential. The paper introduces φ(d), the potential of a design d under its optimal controller, and distinguishes this from Fc(d), the performance of the design under the current shared controller πc. However, is defined only in terms of Fc(d), leaving φ(d) out of the generalization metric. As a result, ρc cannot distinguish between a controller that fails to generalize to d and a d that is simply a bad design.

Controller generalization appears to closely track designer spread Table 7, and the authors note that “a broader body distribution for the designer to sample evidently gives the controller more to generalize over, independent of the controllerʼs own entropy”. However, this fact is not reconciled with the claim that SCOPER improves controller generalization through “representation learning and a shared-network design”. Another plausible explanation for variation in controller generalization is that Pk is constructed relative to each designerʼs distribution, not a fixed distribution. Thus, for designers with greater spread, the test bodies (k = 1-10) will be more likely to overlap with the controllerʼs training distribution, so ρc partly measures in-distribution performance, not generalization. Moreover, since Pk is constructed separately or each method, comparing ρc between methods seems questionable to begin with. Similarly, designer spread is measured on different objects in each method Table 5, raising more doubt that the metrics are meaningful axes of comparison.

Since NGE evaluates each child after fine-tuning a controller inherited from its parent, the measured ρc (zero-shot performance of the inherited controller on the mutated child) is not the quantity NGEʼs selection actually operates on. Nonetheless, the authors use it to claim NGE controllers “over-fit and prohibit proper design-space search.” Further, ϵd and ρd are reported as undefined for NGE because it lacks a designer critic; but NGE as published includes one GMUC, a GNN fitness predictor), which the authors disabled. Broadly, the metrics are only fully applicable to designers trained with a critic.

In addition, the empirical superiority of the proposed method is diminished by its sample inefficiency. SCOPER requires 2x the number of environment samples over BodyGen to achieve similar return Figure 4. The specialized return of SCOPER and BodyGen being approximately equal at 25M samples Figure 4, stars) provides another indication that SCOPER is not superior to BodyGen at lower sample counts. It's unclear whether gains in wall-clock time reflect engineering efforts primarily applicable to SCOPER.

Questions for authors

Can ρc be reported on a fixed, method-independent set of bodies for each k?

Why is GMUC not included the NGE baseline? Their manuscript explicitly states it was included to “balance between exploration and exploitation”, which appears to be a primary aim of your study. Also, with it in place, it should be possible to evaluate designer exploration and generalization.

Is a specialized return score for SCOPER available at BodyGenʼs 50M sample budget?

What do the robots look like that were designed by SCOPER?

4 comments
AuthorsOpenReview ↗

Reviewer Gi2h

Thank you for the detailed review, and for recognising that how morphology and policy optimisers should differ is an important, understudied question, and that the commitment gap is a useful lens on it.

@1) ρC\rho_C, fixed-body evaluation, and controller competence

As in Common Response C1, we agree that ρC\rho_C does not isolate morphology-independent controller ability. We now call it controller competence transfer: how well realised performance is retained over the morphologies each method's own designer produces, not unseen-body generalisation.

To answer your question directly, we evaluate all frozen controllers on a fixed bank of 56 morphologies, normalised by body-specific specialists. The ρC\rho_C ranking does not persist: BodyGen achieves the highest normalised shaped return, SCOPER substantially greater forward progress. The two answer different questions, so we report the common bank as a cross-method check and keep ρC\rho_C as a method-local diagnostic.

Your observation that designer spread can itself raise ρC\rho_C is important. In our entropy sweep, higher generator entropy broadens training morphologies and coincides with higher competence transfer, though the sweep does not isolate exposure as the cause. Exposure alone is insufficient, however: SCOPER's split and decoupled ablations keep much broader designer spread but perform far worse, so broad search must be paired with coordinated control.

@2) NGE adaptation and GM-UC

The NGE diagnostic is measured immediately after mutation, before the inherited controller adapts to the child, so it measures immediate competence transfer, not selection fitness. It does not show that NGE's adaptation budget is insufficient, only that adaptation must recover a substantial gap; if recovery is incomplete, selection can favour quickly-adapting designs over higher-potential ones. We clarify this in Sec. 4.2.

Our baseline used the released default, with GM-UC disabled. With GM-UC enabled, at generation 80 (~215.3215.3M transitions), frozen-controller retention is 0.2642±0.02030.2642\pm0.0203 vs. 0.2597±0.03070.2597\pm0.0307 without, and population return 3997.1±336.43997.1\pm336.4 vs. 3930.6±262.93930.6\pm262.9; best-so-far scores at generation 70 are 5567.25567.2 vs. 5580.95580.9. GM-UC therefore neither removes the post-mutation gap nor consistently improves performance. Since it provides the morphology predictor needed for designer-side diagnostics, we use NGE+GM-UC for those, and state which configuration each result uses.

@3) Sample efficiency and the 50M comparison

SCOPER is not presented as more sample-efficient than BodyGen. It uses more interactions (201201M vs. 5050M) but reaches higher return (74417441 vs. 49714971) in less wall-clock time (7.37.3 h vs. 17.617.6 h), partly through parallelism (Table 2). Its role is to test whether the controlled study's principles carry over to robotic co-design.

At BodyGen's 50M budget, using the same specialisation protocol, SCOPER has already found a morphology whose specialised performance is greater than BodyGen's final one, while continuing to search beyond it. We do not claim superior sample efficiency from this. We believe the larger discrepancy between SCOPER's un-specialised reward and its specialised reward is due to having a higher spread of samples, which is difficult for the controller to be fully competent across, and because SCOPER optimises morphology more than the BodyGen, which we demonstrate collapses onto a single design quickly.

@4) Morphology visualisations

We add early, middle, and final committed morphologies for all three methods (App. D, Fig. 6).

Controlled-study vs. robotics metrics

This difference is deliberate. The toy study sets spread, exploration, and competence transfer directly; in real methods they emerge jointly, so we measure operational diagnostics instead, and do not assume toy results transfer automatically. Likewise, Table 5 measures the distribution each method actually searches (NGE population, BodyGen replay buffer, SCOPER samples), so spread characterises each method's search rather than ranking methods exactly.

You correctly note that controller spread changes both action search and the fitness used to evaluate designs. This coupling is inherent to co-design, since morphology quality is observed through rollouts; Fig. 2 therefore shows the net effect of controller spread, not whether the drop comes from broader action search or noisier morphology evaluation.


We hope these evaluations and clarifications address your soundness concerns. If so, we would appreciate your reconsideration of the rating; otherwise, we would be grateful to know which concerns remain.

I thank the authors for their response and updates to the paper. I retain my review. I have remaining significant concerns:

 

(1) I appreciate that SCOPER achieves higher return than BodyGen in less wall-clock time; however, the authors attribute the wall-clock gap to parallelization. Although they state that BodyGen was parallelized as far as its algorithm permits, the gap then reflects the implementation and simulator as much as the method, and it remains unclear how optimized the BodyGen implementation is. The wall-clock / performance advantage appears to be the main results-based claim of the paper. Quoting the abstract: SCOPER achieves “∼50% higher task performance than prior co-design methods while halving runtime.” The wall-clock claim is confounded by implementation, and the performance claim is confounded by sample count: it compares SCOPER at 201M samples against BodyGen at 50M. Moreover, in the rebuttal the authors write that the purpose of the comparison to BodyGen is “to test whether the controlled study’s principles carry over to robotic co-design.” However, the authors also state in the manuscript:

“The test environment shows what co-design should look like if each property could be set by hand. Real methods offer no such control. Their spread, exploration, and competence transfer emerge together from architecture and parameterization, and none is tuned in isolation.”

If the properties cannot be set independently on robots, then the comparison to BodyGen cannot isolate the effect of any one property. Moreover, the two methods differ greatly in details. This leaves me unsure what the BodyGen comparison is meant to establish and what the related claim is meant to be.

 

(2) Changing to controller competence transfer does not resolve the issue. The authors acknowledge that “equal structural distance does not imply equal morphology potential”, and that controller competence transfer therefore “should not be interpreted as a ranking of morphology-independent controller ability”. Yet, the paper’s introduction states:

“… morphology selection can reflect the capabilities of the controller used for evaluation, rather than morphology potential alone. We study how morphology-search breadth interacts with controller competence transfer, and how limited competence transfer can cause promising morphologies to be lost during selection.”

The metric cannot separate these two effects. Controller competence transfer is the ratio of realized return on nearby designs to realized return on the committed design, with no reference to morphology potential. A low value could mean the controller fails to realize the potential of a good morphology, or simply that the nearby morphologies are worse. Only the former corresponds to a promising morphology being lost.

Moreover, the only evaluation that does normalize by specialist performance for each morphology (Appx. G.2) indicates that SCOPER is not superior.

AuthorsOpenReview ↗

Response to Reviewer Gi2h (1/2)

Thank you for the follow-up.

@1) Role of the BodyGen comparison and SCOPER results

We notice that our earlier wording that the BodyGen comparison “tests whether the controlled study’s principles carry over” was too broad. The BodyGen/NGE comparisons are not intended to isolate the causal effect of any individual property.

Instead, the experiments are deliberately complementary. The controlled study provides causal isolation by varying spread, exploration, and competence transfer independently (Sec. 3.4; Figs. 2--3). The robotics comparisons ask whether the resulting diagnostics reveal informative optimization profiles and potential bottlenecks in real co-design algorithms, where these properties necessarily emerge jointly (Sec. 4.2; Fig. 4). SCOPER’s within-method sweeps and ablations, rather than the cross-method BodyGen comparison, then test whether design choices motivated by this framework affect the targeted diagnostics and co-design performance (Sec. 5.2; Table 1).

Thus, the BodyGen/NGE comparison is a diagnostic case study of the framework in realistic coupled systems, rather than an intervention on any individual metric.

Cross-method performance: We agree that the previous abstract wording could reasonably make the SCOPER--BodyGen performance/runtime comparison appear to be the paper’s main empirical claim. That was not the intended role of this comparison. We therefore remove the “∼50%\sim50\% higher task performance while halving runtime” headline from the abstract and do not interpret the comparison as evidence of greater sample efficiency or an implementation-independent runtime advantage (Sec. 5.3; Table 2).

As described in our Common Response C2, the paper’s central empirical evidence is instead the controlled demonstration of the designer--controller asymmetry and its consequences, together with the robotics diagnostics and within-method experiments showing how these ideas can be used to analyse and construct co-design algorithms.

AuthorsOpenReview ↗

Response to Reviewer Gi2h (2/2)

@2) Controller competence and morphology potential

As detailed in Common Response C1, we agree that ρC\rho_C alone cannot distinguish poorer controller competence from lower morphology potential; renaming it was intended to clarify what it measures, not remove this confound.

To test whether the observed retention loss is explained by differences in body-specific performance, we normalize the frozen controller's return on each evaluated morphology dd by an independently trained specialist on that same body (App. H.3),

qC(d)=FC(d)Fspec(d),q_C(d)=\frac{F_C(d)}{F_{\mathrm{spec}}(d)},

and compare this specialist-normalized competence on nearby morphologies with that on the committed morphology.

MethodCommitted-body competenceNeighbour competenceNormalized retention
SCOPER51.3±4.9%51.3\pm4.9\%41.8±3.4%41.8\pm3.4\%81.9±10.9%81.9\pm10.9\%
BodyGen78.0±5.0%78.0\pm5.0\%41.4±13.7%41.4\pm13.7\%52.6±15.0%52.6\pm15.0\%
NGE65.6±14.3%65.6\pm14.3\%26.4±4.2%26.4\pm4.2\%42.4±15.9%42.4\pm15.9\%
NGE + GM-UC70.7±8.9%70.7\pm8.9\%25.0±5.6%25.0\pm5.6\%35.8±8.9%35.8\pm8.9\%

The local competence drop therefore persists after accounting for body-specific specialist performance (Table 6; Fig. 14), so poorer neighbouring morphologies alone do not explain the observed retention loss. The SCOPER--BodyGen comparison makes this particularly clear: BodyGen achieves substantially greater competence on its committed body, yet the two achieve almost identical competence on nearby bodies, producing markedly different retention. SCOPER’s higher retention therefore reflects a smaller relative drop from its committed-body performance, not higher absolute competence on the neighbouring morphologies. We treat these finite-budget specialists as empirical references rather than true morphology potential φ(d)\varphi(d).

The stronger selection claim is established separately. The finding that higher-potential morphologies can be discovered but lost during selection comes from the controlled commitment-gap experiment in Sec. 3.4, Fig. 3, where φ(d)\varphi(d) is known directly. Increasing designer spread discovers higher-potential morphologies, while limited controller competence can widen the gap between the best potential encountered and the morphology ultimately selected. We therefore revise the quoted introduction passage to locate this claim explicitly in the controlled setting, rather than infer it from robot-side ρC\rho_C.

Why retain ρC\rho_C on robots? Although ρC\rho_C alone cannot attribute a return drop to the controller rather than the morphology, it is a search-conditioned diagnostic of the realized-return landscape on which co-design actually operates. The controlled analysis shows that controller shortfall matters over the designs produced by the designer’s search (Sec. 3.4; Eq. 7): for broader search to be useful, the controller must evaluate the resulting morphologies reliably. Our PkP_k probes therefore use designer-induced perturbations around each method’s committed morphology to measure local retention at controlled structural distance (Sec. 4.1). This does not remove the morphology-potential confound, but makes the diagnostic directly relevant to the morphology region over which competence is required during co-design.

It is also practical to obtain directly from a co-design run, whereas specialist normalization requires independently training controllers for each evaluated morphology. We therefore view ρC\rho_C as a low-cost operational diagnostic, with specialist-normalized evaluation providing a substantially more expensive check on its interpretation (App. H.3).

The common-bank evaluation in App. H.2 asks a different, complementary question: how the frozen controllers transfer over the same broad, method-independent set of 56 morphologies. This bank is itself necessarily sparse relative to the approximately 3.69×10173.69\times10^{17} possible designs, and is intended as a structured probe of broader transfer rather than an exhaustive measure of controller ability (Fig. 11). It does not establish uniform SCOPER superiority, and we do not claim that it does. Moreover, it yields no single consistent ranking: BodyGen achieves higher specialist-normalized shaped return, whereas SCOPER achieves greater signed forward progress, and the zero-action baseline shows that auxiliary reward terms materially contribute to shaped return on this bank (Fig. 12).

We therefore use the common bank to probe broader method-independent transfer, while ρC\rho_C characterizes the search-conditioned retention relevant to each method’s own co-design process, with the local specialist-normalized evaluation providing the more direct check on the morphology-quality confound in ρC\rho_C.

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

The paper studies how morphology designers and controllers benefit differently from sampling breadth, exploration, and generalization. It introduces operational metrics and a controlled synthetic landscape, arguing that designers benefit from broad search while controllers benefit more from generalization and relatively greedy actions. An expected-value-of-perfect-information decomposition motivates this asymmetry. The proposed SCOPER combines asymmetric entropy bonuses, a shared Transformer, physical auxiliary losses, and frequent morphology resampling. On Modular Ant, it reports a return of 7441 versus BodyGen's 4971, with higher sample use but lower wall-clock time, supported by representation and entropy ablations.

Critical strengths
  1. Separating a morphology's potential, its performance under the current controller, and the designer critic's estimate helps explain why a promising body can be discarded. The commitment-gap experiment makes this distinction concrete: broader search can discover better bodies without improving the final selected body. This is a valuable diagnostic perspective beyond comparing final task returns.

  2. SCOPER's shared-representation and decoupling ablations show that broad morphology sampling alone is insufficient: the split and decoupled variants retain broader designer distributions yet perform worse. Eight-seed ablations, specialized-policy evaluations, entropy sweeps, and both wall-clock and sample-count plots provide useful evidence for the practical combination of representation sharing and sustained search breadth. The paper also acknowledges that the entropy controls affect more than their intended metric.

Critical weaknesses
  1. Equation (5) defines controller generalization as the return on perturbed bodies divided by the return on the committed body. Even an optimal controller on every body would give E[φ(d)] / φ(d̄), which need not equal one. Thus, a drop can reflect intrinsically weaker bodies, rather than failed controller transfer. This matters because each method constructs its own perturbation distribution by widening its designer, so equal edit distance does not imply comparable bodies or potential. The specialized evaluation in Appendix F.0.1 is helpful, but fine-tunes each method's existing controller on its own committed body; it does not remove differences in architecture, initialization, or optimization history.

  2. Equation (6) correctly shows the value of adapting an action after observing the design, relative to using one fixed action across designs. It neither orders the designer's sampling cost against the controller's nor establishes that increasing design spread improves performance. The supporting toy experiment uses one-step rewards, fixed-variance sampling, and local gradient ascent as a proxy for generalization, leaving out the information-gathering role of actions in sequential RL. Appendix E itself reports a reversal of the designer-spread effect at large dimension and sampling ratio. The robot evidence is confined to one Modular Ant task and grammar, as the limitations acknowledge. These results support a useful conditional heuristic, but do not yet justify the general conclusion that controllers should sample greedily while designers should maintain breadth.

Questions for authors
  1. Does the controller-generalization ranking persist on a common set of perturbed morphologies, with returns normalized against independently trained specialist controllers for those same bodies? Even a representative subset would help determine how much of the measured decline reflects body potential rather than transfer failure.
  2. Which assumptions, beyond the nonnegative EVPI identity, justify the proposed relative preferences for sampling breadth? Please distinguish conclusions established by Eq. (6) from observations specific to the toy optimizer, and explain how the reversal in Appendix E limits the recommendation for practical co-design.
4 comments
AuthorsOpenReview ↗

Reviewer ZPXN Response

Thank you for the insightful review. We appreciate your recognition that the commitment-gap analysis provides a useful diagnostic beyond final return, and that the SCOPER ablations show broad morphology sampling alone is insufficient.

1. Common-body evaluation and Eq. 5

As detailed in Common Response C1, we agree that ρC\rho_C does not isolate controller competence from morphology potential, and that method-local PkP_k distributions prevent its absolute values from being interpreted as a controlled cross-method ranking.

We therefore performed the suggested common-body evaluation: all frozen controllers are evaluated without adaptation on the same 56 morphologies using independently trained body-specific specialists as common references. The original ρC\rho_C ranking does not persist: BodyGen achieves the highest specialist-normalised shaped return, whereas SCOPER achieves substantially greater signed forward travel. The zero-action control further shows that substantial shaped return from auxiliary rewards can arise without locomotion.

We now interpret ρC\rho_C as method-local controller competence transfer, rather than morphology-independent controller superiority, and add the full common-bank evaluation in App. G.2, Figs.~9--10.

2. Eq. (6), sampling breadth, and Appendix E

Eq.~(6) establishes the structural asymmetry underlying the sampling-breadth preference; under a fixed search budget it suggests that preference, which the controlled experiments then test. Greater spread broadens the distribution over possible designs or actions, increasing sampling uncertainty. On the designer side, that uncertainty is resolved before the controller acts, so the controller can condition its response on the realised design; the EVPI term captures the value of this downstream response and partially offsets the cost that design uncertainty would otherwise impose. No analogous downstream decision is available after a controller action is executed. Thus, Eq. (6) explains why designer-side sampling uncertainty can be more recoverable, but does not determine how much spread is beneficial.

The controlled toy study then provides empirical evidence for the relative spread preference. Sweeping spread independently shows that greater designer spread generally improves best-design fitness, whereas greater controller spread reduces it (Fig. 2). Eq. (7) additionally accounts for the fact that the controller available at that point in optimisation may not realise the optimal design-conditioned response assumed by Eq.~(6). The controller-competence and commitment-gap experiments show the consequence: broader designer spread can discover higher-potential designs, but if controller competence does not transfer across them, their realised fitness can be underestimated and they may be lost during selection; stronger controller competence transfer narrows this gap (Fig. 3). Together, these results identify two conditions for broader designer sampling: (1) it must provide access to higher-potential designs, and (2) the controller must remain sufficiently competent across them for their value to be recognised.

The reversal at N=16,κ=256N=16,\kappa=256 identifies a third condition and directly limits the recommendation for practical co-design. Under the fixed evaluation budget, large κ\kappa allocates many controller updates to each design batch, leaving few morphology-search updates; in this regime, the designer-spread effect reverses (App. F, Fig. 7; App. E in the original submission). Broader designer sampling is therefore useful only when (1) it exposes higher-potential designs, (2) controller competence is sufficient to evaluate them, and (3) enough morphology-search budget remains to exploit those discoveries. When controller optimisation consumes most of the available budget, a more conservative morphology search can instead be preferable.

We have revised Sec. 3.4 to distinguish the structural result established by Eq. (6) from the empirical spread preference, and to state that EVPI is a one-step quantity, which the N=16,κ=256N=16,\kappa=256 reversal illustrates.


We hope these additional evaluations and revisions address your concerns, and would appreciate your reconsideration of the paper's rating if so. We would also be grateful for any remaining concerns we can clarify during the discussion.

Thanks for the reply. That clears up some of my concerns, but my main one is still open. The abstract says SCOPER uses representation learning "to increase controller competence transfer". The Table 1 ablations don't seem to support that, and none of them were run on the common morphology set, so I still don't see evidence for it. I agree with Reviewer Gi2h on this point.

AuthorsOpenReview ↗

Response to Reviewer ZPXN (1/2)

Thank you for following up. We agree with the specific concern: the representation ablations do not establish that sharing the designer and controller backbone itself increases controller competence transfer. We have therefore removed this attribution from the abstract and manuscript (Abstract; Secs. 5.2 and 6).

This distinction is also consistent with the ablations. SCOPER uses a BodyGen-inspired morphology-conditioned Transformer controller, motivated in part by BodyGen's relatively strong local controller retention (Secs. 4.2 and 5), and the representation variants retain this morphology-conditioned control setting (App. K). Their local retention therefore remains relatively high and does not vary consistently with representation sharing. Similarly, removing the FD/FK physical auxiliary objectives has little effect on local retention (0.97±0.040.97\pm0.04 versus 0.98±0.020.98\pm0.02 for SCOPER), suggesting that these representation-learning objectives are also not the primary source of controller competence transfer (Table 1).

We additionally ran the representation variants on the same 56-body common morphology bank (Sec. H.2; Fig. 13). This confirms that representation sharing does not produce a consistent advantage in broader controller transfer; in particular, Split retains greater signed forward travel at later checkpoints (Fig. 13). This is consistent with Split maintaining a substantially broader designer distribution during training (Fig. 21), and our entropy sweep independently shows that broader designer sampling can increase controller competence transfer through increased morphology exposure (Sec. 5.2; Fig. 19; Table 8). We therefore no longer attribute SCOPER's controller competence transfer to the shared backbone.

The representation ablation nevertheless has a substantial effect on co-design performance. Separating the designer and controller trunks in Split reduces return by 37%37\%, and removing the remaining controller-to-designer distillation in Decoupled reduces it by 61%61\%, despite both variants maintaining substantially broader designer distributions (Sec. 5.2; Table 1; Fig. 21). Thus, broader morphology exposure and broader controller transfer alone do not explain successful co-design. The question is instead how morphology search and controller learning remain coordinated.

Our revised interpretation is primarily designer-side. The controlled study shows that morphology-search breadth and controller competence must remain coordinated: the controller must remain useful over the region the designer searches (Sec. 3.4; Fig. 3). Moreover, the designer does not optimize the unknown morphology potential φ(d)\varphi(d) directly; it learns the moving controller-dependent fitness landscape VD(d)≈FC(d)V_D(d)\approx F_C(d) (Sec. 3.1). SCOPER therefore shares the designer and controller representation so that information learned during controller optimization can directly update the representation used for morphology-value prediction and hence influence morphology search (Sec. 5).

We added diagnostics targeted at this interpretation. On the morphology batches actually encountered during training, SCOPER predicts subsequently realized controller return with substantially lower normalized error than Split across four matched budgets (0.920.92--0.980.98 versus 1.591.59--2.242.24) and with markedly smaller bias (Sec. 5.2; Figs. 20--22; Table 9). This advantage is already present at the earliest checkpoint, while both designers still sample broadly (Fig. 21). These results support the interpretation that sharing control-learned features helps the designer track the changing controller-dependent fitness landscape.

We also evaluate competence over each variant's actual designer distribution. Across the four matched budgets, Split retains 0.961/0.751/0.868/0.8040.961/0.751/0.868/0.804 of committed-body return, while SCOPER remains approximately at full retention (Fig. 22). At the final checkpoint, Split samples morphologies an average of 4.104.10 modules from its committed body and retains only 0.804±0.0110.804\pm0.011 of committed-body return over these samples (Fig. 22; Table 11). This is consistent with the controlled-study result that broad search must be matched by sufficient controller performance over the region being searched (Fig. 3). Because realized return also depends on morphology quality, we do not interpret this as an isolated representation-sharing →\rightarrow controller-transfer effect.

AuthorsOpenReview ↗

Response to Reviewer ZPXN (2/2)

Taken together, these results suggest a distinction between broad transfer and co-design coordination. Split's stronger transfer on parts of the fixed morphology bank is consistent with its broader training distribution and the increased morphology exposure that this provides (Figs. 13 and 21; Fig. 19 and Table 8 provide the complementary exposure ablation), while this does not by itself ensure that the designer accurately tracks or exploits the controller-dependent landscape over the morphologies it is searching. Representation sharing substantially improves co-design performance and is associated with better designer-side prediction and calibration of that landscape (Table 1; Figs. 20--22; Table 9), while the evidence does not support a consistent improvement in broader controller competence transfer (Fig. 13; Table 10).

Comments on the paper

4 comments
AuthorsOpenReview ↗

Reviewer Response

We thank all our reviewers for your constructive feedback, and your time. We have thought hard about your concerns and suggestions, and have made updates to this paper which we hope will have addressed your concerns and made confusion portions more clear.

We have made many changes to the paper, the important ones which are written in red for you to see. Below is a more direct response to common concerns, in addition to a response that each reviewer will recieve catered to their specific concerns.

If you find that these clarifications and adaptations assuage your concerns, we ask you to increase your rating of categories which have improved meaninfully.

Thank you

Common responses:

C1: Controller generalisation (ρC\rho_C) concerns

1. What does ρC\rho_C measure?

We thank the reviewers for highlighting this distinction. We agree that ρC\rho_C does not isolate morphology-independent controller ability: because it uses realised return FC(d)F_C(d), lower values can reflect poorer controller transfer, lower morphology potential ϕ(d)\phi(d), or both.

We therefore refer to ρC\rho_C as controller competence transfer. It measures how well the current controller's realised task competence is retained as morphology changes away from the committed design. Thus, ρC=1\rho_C=1 means realised return is retained, not necessarily that morphology-independent controller ability is unchanged.

We make this explicit in Sec.3.2, Eq. (5), building on the distinction between morphology potential ϕ(d)\phi(d) and current-controller fitness FC(d)F_C(d) in Sec. 3.1.

2. Why is PkP_k local?

PkP_k is intentionally defined over the morphology region explored by each method. In co-design, the controller need not perform uniformly over the entire morphology space; it must remain competent over the morphologies produced by its own designer. We therefore construct PkP_k from each method's search distribution: for SCOPER and BodyGen through their designer distributions, and for NGE through morphology mutations (App.G.1, Table 5).

This makes ρC\rho_C directly relevant to the co-design process, but limits its cross-method interpretation: equal structural distance does not imply equal evaluated morphologies or equal morphology potential. We therefore use ρC\rho_C as a method-local measure of competence transfer, rather than an apples-to-apples ranking of morphology-independent controller ability (Sec.4.1).

3. Improve cross-method comparison

To attempt to compare across methods more fairly, we performed the requested common-body evaluation. All frozen controllers are evaluated without adaptation on the same 56 morphologies, selected independently of controller performance and stratified across three reference bodies, structural distances (0,2,4,8,12)(0,2,4,8,12), and reachable leg counts from one to eight. Independently trained specialists provide the same body-specific reference for every method (App. G.2; Figs. 9--10).

Method / referenceReturn (% specialist reference)Forward progress (% specialist reference)
Zero action11.8611.86−0.0035-0.0035
SCOPER7.13±1.037.13 \pm 1.031.233±0.1971.233 \pm 0.197
BodyGen archive10.83±0.3010.83 \pm 0.30−0.023±0.015-0.023 \pm 0.015
NGE baseline3.69±4.093.69 \pm 4.090.174±0.2950.174 \pm 0.295
NGE + GMUC5.70±2.505.70 \pm 2.500.358±0.2440.358 \pm 0.244

The common-bank evaluation does not reproduce the original ρC\rho_C ranking: among learned controllers, BodyGen achieves the highest specialist-normalised shaped return, whereas SCOPER achieves substantially greater signed forward travel. However, zero action variant itself obtains 11.8611.86% of specialist shaped return while producing essentially no forward movement - in other words, all methods effectively failed to move at all, and standing was the best they could hope for. BodyGen's shaped-return result is comparable to this zero-action reference and is likewise accompanied by near-zero signed forward travel. This demonstrates that the controllers of these co-design methods are locally optimal, but are not globally so, as we would expect.

These results clarify that the two evaluations answer different questions. ρC\rho_C measures competence transfer over the morphology region induced by each method's own search, whereas the common bank evaluates broader transfer on identical bodies. Common-bank shaped-return comparisons can also be confounded by auxiliary reward terms, particularly when they favour behaviours that compete with the main task objective. To emphasize this distiction, we assert that ρC\rho_C is not a ranking of morphology-independent controller ability, and report the common-bank envalutations as complementary diagnostics. This is made explicit in Sections 3.2 and 4.1, with full evaluation details including common-bank performance trajectories across training checkpoints in Appendix G.2.

Dear reviewers,

As we only have a few more days to go before the discussion window closes, please go through the author responses and acknowledge them. You may highlight aspects of your initial review comments that have been satisfactorily addressed by the responses, as well as those which require further clarification.

Thanks,

AC

Common Response 2: Clarifying the paper's contribution

We thank the reviewers for helping us sharpen the scope and interpretation of the paper. Our central contribution is not a state-of-the-art cross-method comparison, but new understanding of how the two optimizers in robot co-design benefit from different optimization behaviour, why this asymmetry arises, and how these insights can be used to analyse and design co-design algorithms.

More specifically, we ask whether the designer and controller should be optimized differently despite sharing the same final objective. We introduce three diagnostics, spread, exploration, and competence transfer, and construct a controlled environment in which these properties can be varied independently. This reveals a systematic asymmetry: broader designer spread generally improves discovery, whereas controller spread is detrimental, and controller competence transfer benefits the controller more strongly than the designer. We relate this asymmetry to the sequential structure of co-design through the Expected Value of Perfect Information (Sec. 3.4; Eq. 6; Fig. 2). The controlled commitment-gap experiment further shows, in a setting where morphology potential is known directly, that broad search can discover higher-potential designs while limited controller competence can cause some of their value to be lost during selection (Fig. 3).

The robotics experiments play a complementary role. In real co-design systems, spread, exploration, and competence transfer emerge jointly from the architecture and training procedure, so these experiments do not isolate the causal effect of individual properties. Instead, they ask whether our diagnostics reveal meaningful optimization profiles and potential bottlenecks in realistic algorithms. NGE and BodyGen exhibit contrasting profiles: NGE maintains broader morphology search but has a larger immediate transfer gap following morphology changes, whereas BodyGen retains stronger local controller performance but concentrates its morphology search early (Sec. 4.2; Fig. 4).

We then use these observations to construct SCOPER. SCOPER combines broad designer sampling, low controller sampling spread, morphology-conditioned Transformer control, and shared design--control representations (Sec. 5; Fig. 5). We build on BodyGen's morphology-conditioned Transformer controller, which exhibits relatively strong local controller retention in our diagnostics, and train it continually across the changing morphologies produced during search. Separately, the shared representation allows information learned through controller optimization to directly inform the designer's model of controller-dependent fitness, coupling controller learning to morphology search. This is motivated by the fact that the designer does not optimize the unknown morphology potential directly, but instead learns the current controller-dependent fitness landscape, VD(d)≈FC(d)V_D(d)\approx F_C(d) (Sec. 3.1). Our ablations show that representation sharing has a substantial effect on co-design performance (Table 1) and, in additional diagnostics, is associated with substantially better prediction and calibration of controller-dependent fitness on the morphologies encountered during training (Figs. 20--21; Table 9).

Importantly, we do not claim that representation sharing itself increases controller competence transfer. The representation variants retain relatively high local controller retention, and the common-morphology evaluation does not show a consistent shared-backbone advantage (Fig. 13; Table 10). We therefore distinguish the controller-side competence required when the controller is reused across changing morphologies from the designer-side benefit of allowing controller learning to inform the morphology search through the shared representation.

The paper's primary contribution is therefore the identification and explanation of the designer--controller asymmetry, together with diagnostics that reveal how this asymmetry manifests in realistic co-design algorithms. SCOPER is an example of using these insights constructively, rather than the primary basis for a state-of-the-art claim.

AuthorsOpenReview ↗

Concluding Notes

We want to thank all of our reviewers for their time, consideration, and construtive feedback. We are pleased to hear that you find our metrics and commitment gap to have a "valuable diagnostic perspective", and that it "offer(s) a meaningful perspective beyond developing another co-design algorithm". Based on your feedback, we've improved the paper's clarity around our metrics, and run several additional experiments which re-affirm our previous findings. We believe that thanks to your feedback, this paper is both much clearer and well evidenced, and that our methods can meaningfully contribute to future co-design endeavors.

New Experiments Run

1. Common Body Bank

The most common request we had was to demonstrate how each method generalised over a prechosen set of bodies. We chose not to perform this experiment originally because any set chosen would be biased, and more importantly, because these co-design methods all optimise relatively locally. This experiment demonstrated these issues as expected, as well as issues regarding reward hacking, as some methods performed well on unseed morphologies by standing still without trying to move. Notably though, this did confirm that SCOPER had the widest effective controller competence transfer of the entire morphology space, if you consider forward travel instead of raw reward (although all methods covered relatively little).

2. Two more tasks

To give evidence that SCOPER performs across a variety of tasks, we trained on both the Gap task (a locomotion task), and Relocate (a grasping task). We found that SCOPER was able to complete both tasks without major hyperparameter tuning (only gamma), and show that its metrics over training hold the same patterns as previously demonstrated on Ant.

3. Further Ablation Analysis

One reviewer noted that there was still some ambiguity as to how much controller competence was improved by the methods tested in our ablation. We analyzed our ablations more fully, and found that while no individual modification was a major contributor to controller competence, the shared-backbone architecture substantially improved co-design performance and was associated with more accurate designer prediction and calibration throughout training.

Paper Modifications

We have denotes meaningful changes in the pdf in red. Below is a summary of the modifications to the paper.

1. Generalization becomes Competence Transfer

Based on insightful feedback from our reviwers, we concluded that the name 'competence transfer' for our third metric was more accurate and clear.

2. Qualitative Visual Comparison

Although our website showed visual comparison between the designs produced by our methods, we neglected to add the same to the paper, and have provided a diagram showing how bodies evolve for each method over time, alongside a qualitative analysis explaining how we interpret these morphologies relative to our findings.

3. Improved Introduction

Our introduction was not clear about the overall contribution of the paper, and gave the false impression that our primary contribution was providing a new co-design algorithm. We have updated this to be more accurate and clear.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.