acceptodds
Under review as a conference paper at ICLR 2027

Bridging Body and Brain: Gene-Driven Morphology–Control Co-Design

Abstract

Morphology–control co-design jointly optimizes an agent’s body structure and control policy as an integrated embodied system. However, existing methods typically model morphology design and control with separate networks coupled only indirectly through a shared task objective, limiting explicit high-level coordination. Inspired by natural genes that coordinate biological development, we introduce Morphogene, a compact latent blueprint that bridges an agent's body and brain. Through AdaConcat, Morphogene jointly conditions morphology and control generation at the limb level, allowing its variations to induce coordinated changes in both components. Building on this representation, we propose GeCode, which formulates co-design as exploration in the compact Morphogene space. Each Morphogene anchors a local design region in which nearby body–brain designs are explored, while performance-guided updates move these anchors toward promising regions for more efficient exploration of the broader design space. This process combines local refinement with global exploration while preserving body–brain compatibility. Extensive experiments across diverse 2D and 3D co-design tasks demonstrate that GeCode consistently outperforms existing state-of-the-art methods, achieving substantially faster convergence and higher final performance. The code is available at https://anonymous.4open.science/w/GeCode-BC8B/.

Primary area:
reinforcement learning
Keywords:
Morphology-Control Co-Design, Embodied Co-Design, Latent-space exploration, Reinforcement Learning

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_Hc1FOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2(−1)
Soundness
2
Read review
Paper summary

A compressed latent genome was learned for kinematic trees, the majority of which were planar bodies but in one case the branching limbs extending from a quadrupedal ("Ant") base were evolved. In each case, the latent genotype vector was used for morphology-conditioned control... thus bridging body and brain. Populations of agents were expressed about different points in latent space, the locations of which were updated locally with a bias toward the latent coordinates of the very best (globally) across all subpopulations.

Critical strengths

Potentially the state-of-the-art for this particular benchmark, which is related to robot design.

Critical weaknesses

There are two critical weaknesses. The first is poor scholarship. The concept of latent design genomes and latent genes in robots is not new [1-3] and conditioning sensorimotor control on the latent genotype has also been demonstrated [3,4]. The basic analogy is outlined in [5]. These references are suspiciously missing from the paper. The novelty of the paper is unclear and hard to appreciate because the contribution is not situated in the literature.

The second weakness concerns the meaning of this benchmark. Yes, the authors might have achieved state-of-the-art results (see questions). But is the art meaningful? The bodies are so strong and light that they literally float across the screen. They have similar density to that of a balloon animal with the torque of a whole car in every single joint. Some context is needed here, and it is needed in the main text. Similar designs have been realized as physical robots [3], but at much lower energy density.

References:

  1. Hu et al. (2023) CoRL. https://proceedings.mlr.press/v205/hu23c.html
  2. Li et al. (2025) ICLR. https://openreview.net/forum?id=awvJBtB2op
  3. Yu et al. (2026) PNAS. https://www.pnas.org/doi/10.1073/pnas.2519129123
  4. Wang et al. (2026) ICML. https://openreview.net/forum?id=dTIxiQ0jhu
  5. Mitchell & Cheney (2025) Trends in Genetics. https://www.cell.com/trends/genetics/fulltext/S0168-9525(25)00008-3
Questions for authors

How does this work differ from [1-4]?

What do the results in Fig. 4 look like when the horizontal axis is wall-clock time instead of env steps?

How do the global best and the local best affect search? Can you isolate the effects with a simple ablation?

How do changes in latent genotypes affect the body plan and behavior (i.e. the phenotype)?

How is topology/connectivity represented in the encoder?

18 comments
AuthorsOpenReview ↗

Response to Reviewer Hc1F (1/4)

Dear Reviewer Hc1F:

We sincerely thank you for your careful review and constructive feedback, and for recognizing GeCode’s strong performance on the robot co-design benchmark.

We address each of your concerns and questions below. The first anonymous link provides the training curves for the tabulated results (Hc1F_Fig1–Hc1F_Fig5), and the second anonymous link provides the supplementary video (Hc1F_Video).

🖼️ Anonymous Link for Figures
👉 https://anonymous.4open.science/r/GeCode-BC8B/rebuttal/Re_Hc1F.pdf

🎬 Anonymous Link for Video
👉 https://anonymous.4open.science/r/GeCode-BC8B/rebuttal/Hc1F_Video.html


Weakness 1 / Question 1: Poor Scholarship.

We thank the reviewer for bringing these closely related studies to our attention. References [1–4] adopt a common strategy: they encode the high-dimensional, discrete morphology space into a lower-dimensional continuous latent space and perform design optimization there, typically using evolutionary or Bayesian optimization.

As you note, morphology search in a compact, continuous latent space learned by a VAE is well established. GeCode’s novelty lies in how it explores this space. This difference in search strategy distinguishes Morphogenes from conventional genotypes and clarifies their role in morphology–control co-design.

Our key distinctions are summarized below.

(1) Direct Latent Decoding vs. Morphogene-Guided Autoregressive Construction

References [2–4] search over individual latent codes and use a fixed pretrained decoder to map each code to a candidate body. In our MuJoCo benchmark for articulated rigid-body design, unconstrained decoding can produce invalid assemblies with infeasible joint attachments or unintended limb interpenetration. Applying this approach would therefore entail substantial engineering effort, either to incorporate hand-specified geometric and kinematic constraints into the decoder or to implement a separate stage for rejecting or repairing invalid bodies before simulation. The separate evolutionary or Bayesian search adds further computational cost through repeated candidate evaluations.

Following BodyGen, GeCode learns morphology construction and control within a unified reinforcement learning framework. The design policy constructs bodies autoregressively, selecting each structural edit, such as adding or removing a limb, based on the morphology produced so far. Each edit preserves the validity of the intermediate design. Learning which edits to apply improves search efficiency, but stepwise exploration can concentrate around familiar structures and miss structurally distinct designs with higher returns.

GeCode broadens morphology search by conditioning the design policy on Morphogenes that anchor distinct regions of the latent morphology space. This combines global exploration across regions with local refinement through RL. All Morphogene updates reuse rollouts collected for policy training, so anchor refinement stays within the existing environment interaction budget.

(2) Instance-Level Representation vs. Prototype-Level Representation

The difference in how GeCode and [1–4] explore the latent morphology space gives Morphogenes a distinct role from that of conventional genotypes. In [1–4], each sampled genotype serves as an instance-level representation. A fixed pretrained decoder maps it to the body representation of a specific candidate design. Optimizing these genotypes therefore amounts to searching over individual robot designs in a lower-dimensional, continuous space that makes the search more tractable.

In GeCode, a Morphogene serves as a prototype-level representation, analogous to a prototype in meta-learning. It captures the shared structural characteristics of a family of related morphologies and anchors a local region of the latent space. GeCode updates these anchors to explore a wider range of latent regions efficiently, while the RL-trained morphology-design policy uses each Morphogene to guide the local exploration of related body variants.

Response to Weakness 1 / Question 1 continued in the next part (2/4).

AuthorsOpenReview ↗

Response to Reviewer Hc1F (2/4)

Weakness 1 / Question 1: Poor Scholarship (Continued).

(3) Body First, then Brain Adaptation vs. Gene-Driven Body–Brain Co-Adaptation

At the level of optimization dependency, latent-space co-design methods such as [2,4] follow a body-first, brain-second process. Although they begin by sampling a genotype, a fixed VAE decoder maps each sampled genotype to a specific morphology. Because the decoder remains frozen during co-design, morphology generation cannot adapt in response to controller learning. In [4], the same genotype is subsequently provided to the MoE gate to improve control of the corresponding body, but the method still follows a body-first, brain-second dependency.

In contrast, selecting a Morphogene does not determine a particular body. Instead, it simultaneously conditions the stochastic morphology generator and the controller, which are jointly trained through a coupled reinforcement-learning process. Rollout feedback jointly refines both policies and the shared Morphogene, enabling the morphology generator and controller to co-adapt within the same design family.


Following your suggestion, we revised the Related Work section to incorporate references [1–5] and clarify how GeCode builds on and differs from prior work. We also broadened our literature review and added related studies [6–9] to place GeCode in the broader context of morphology–control co-design.

References:

[1] Jiaheng Hu, Julian Whitman, and Howie Choset. Glso: Grammar-guided latent space optimization for sample-efficient robot design automation. In CoRL'23.

[2] Muhan Li, Lingji Kong, and Sam Kriegman. Generating freeform endoskeletal robots. In ICLR'25.

[3] Chen Yu, David Matthews, Jingxian Wang, Jing Gu, Douglas Blackiston, Michael Rubenstein, and Sam Kriegman. Agile legged locomotion in reconfigurable modular robots. PNAS 2026.

[4] Yibin Wang, Muhan Li, Zihan Guo, and Sam Kriegman. ECo-MoE: Embodiment-Conditioned Mixture of Experts Increases the Evolvability of Robots. In ICLR'26.

[5] Kevin J Mitchell and Nick Cheney. The genomic code: The genome instantiates a generative model of the organism. Trends in Genetics 2025.

[6] Sam Kriegman, Nick Cheney, and Josh Bongard. How morphological development can guide evolution. Scientific Reports, 2018

[7] Muhan Li, Lingji Kong, and Sam Kriegman. Generating freeform endoskeletal robots. In ICLR'25.

[8] David Matthews, Andrew Spielberg, Daniela Rus, Sam Kriegman, and Josh Bongard. Efficient automatic design of robots. PNAS 2023.

[9] Luke Strgar, David Matthews, Tyler Hummer, and Sam Kriegman. Evolution and learning in differentiable robots. Arxiv 2024.

AuthorsOpenReview ↗

Response to Reviewer Hc1F (3/4)

Weakness 2: Meaning of Benchmark.

We appreciate the reviewer’s question regarding the significance of this benchmark. The benchmark is not intended to demonstrate that the resulting morphologies can be directly manufactured without additional engineering constraints. Instead, it provides a nature-inspired testbed for studying whether reinforcement learning can jointly learn morphology generation and control through environmental interaction, thereby avoiding the separate population-based evolutionary search used in [2,4].

Transform2Act (ICLR 2022 Oral), BodyGen (ICLR 2025 Spotlight), and StackelbergPPO (ICLR 2026 Poster) were all evaluated on this benchmark and serve as our baselines.


These nature-inspired benchmarks use idealized models to study animal-like body plans and locomotion in simulation. For example, wild cheetahs have been recorded at peak speeds of approximately 25.9 m/s25.9\,\mathrm{m/s}, while our evolved CHEETAH agent reaches approximately 23 m/s23\,\mathrm{m/s} (see Hc1F_Fig1 in anonymous link). Its locomotion speed is therefore comparable to the biological scale that motivates the task.

Following your suggestion, we added this context to the revised manuscript and clarified the purpose and physical scope of the benchmarks (lines 406–410 and 777–804).

In response to your description of the simulated robots as having 'similar density to that of a balloon animal with the torque of a whole car in every single joint', we re-examined the morphology and actuation parameters inherited from Transform2Act and BodyGen and assessed body density and actuator strength separately.

Our measurements show that the visually light motion is not caused by an unrealistically low body density. The evolved CHEETAH has limb densities of 952.8952.8–986.2 kg/m3986.2\,\mathrm{kg/m^3} and a volume-weighted density of 982.0 kg/m3982.0\,\mathrm{kg/m^3}, which is close to that of solid ABS engineering plastic (approximately 1,040 kg/m31{,}040\,\mathrm{kg/m^3}).

The visually light motion primarily results from the relatively large MuJoCo gear coefficient (up to 400), which amplifies actuator torque and enables high joint angular accelerations and rapid body movements despite the robot’s realistic mass and density.

SettingMax gearLimb density range (kg/m3\mathrm{kg/m^3})Volume-weighted density (kg/m3\mathrm{kg/m^3})
Original benchmark400952.8–986.2982.0
Solid ABS Engineering Plastic—Approximately 1,040—

To better reflect the actuation limits of physical robots, we reduce the maximum MuJoCo gear coefficient from 400 to 100 and 50 and retrain GeCode and all baselines at each setting. The results are summarized below, with the corresponding training curves and videos Hc1F_Video are available via the anonymous link https://anonymous.4open.science/r/GeCode-BC8B/rebuttal/Hc1F_Video.html.

Max gearMax. joint torque (N m\mathrm{N\,m})Max. single-joint power (kW\mathrm{kW})Transform2ActBodyGenStackelbergPPOGeCode
Original benchmark400400.0029.299542.1 ± 635.011475.1 ± 489.513246.3 ± 299.919888.7 ± 944.6
Restricted gear100100.001.662311.16 ± 246.853350.65 ± 562.444255.77 ± 788.267956.38 ± 749.69
Restricted gear5050.000.701042.07 ± 39.191013.70 ± 109.911093.37 ± 30.343316.09 ± 478.70

Lowering the gear limit makes locomotion more challenging by reducing the available joint torque and power. GeCode still achieves the highest mean return under both restrictions, with its relative advantage over the strongest baseline increasing from 87.0% at a gear limit of 100 to 203.3% at 50. These results demonstrate that GeCode’s advantage does not depend on excessive actuator strength and becomes more pronounced under stricter actuation constraints.

Question 2: Fig. 4 with Wall-Clock Time.

Following your suggestion, we replotted all learning curves in Fig. 4 using wall-clock time in place of environment steps. The table below summarizes the mean training time and final return across the 12 tasks. The updated learning curves (Hc1F_Fig2) and the full computational cost comparison (Hc1F_Tab1) are available via the anonymous link.

Metric (mean ± SEM across 12 tasks)GeCodeStackelbergPPOBodyGenTransform2Act
Training time (hours)22.952 ± 0.76023.421 ± 0.51420.260 ± 0.46712.460 ± 0.521
Final return15885.39 ± 2400.079140.41 ± 1354.808242.09 ± 1320.335194.98 ± 903.94

GeCode achieves a substantially higher mean final return than StackelbergPPO with comparable training time. Transform2Act trains faster but reaches a considerably lower mean final return. The revised learning curves show that GeCode achieves higher returns under the same wall-clock training budget.

Reviewer_1XGUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GeCode, a morphology–control co-design method built on BodyGen. An offline regularized autoencoder learns a 32-dimensional morphology space using reconstruction, topology, and tree-distance objectives. A small set of latent anchors, called Morphogenes, jointly conditions the design and control policies through limb-specific AdaConcat modulation. A proximity reward encourages local exploration, while rollout returns determine anchor updates, elimination, and resampling. Across twelve MuJoCo tasks and five seeds, GeCode reports higher returns and fewer interactions to reach baseline performance than BodyGen, StackelbergPPO, and Transform2Act.

Critical strengths
  1. Sharing a compact conditioning variable between body generation and control gives the two policies a common description of the intended morphology family. Updating these anchors from existing rollouts is a practical addition to a PPO-based co-design system. The Gene2Body/Gene2Brain ablations support the usefulness of supplying this information to both components, rather than only improving morphology generation.

  2. The twelve settings cover planar and three-dimensional locomotion, terrain traversal, manipulation, and different branching constraints. Improvements are substantial on several tasks: Pusher rises from 3054.4 to 6250.2 relative to StackelbergPPO, and Crawler from 9255.3 to 16143.3 (Table 3). Five-seed results, conditioning ablations, sample-efficiency comparisons, and accounting for offline RAE training make the practical case stronger than a single-task demonstration.

Critical weaknesses
  1. Removing conditioning from either network establishes that both inputs are useful, but does not separate the benefit of a shared searchable anchor from additional morphology features, the proximity reward, and the changed search distribution. Figure 6 evaluates structural regularizers through latent-space metrics, without showing their effect on co-design return; the proximity bonus is also not independently ablated.

  2. Equation (14) applies softmax directly to top-rollout mean returns, without a stated temperature or return normalization. At the reported return scales, even modest differences can make sampling almost deterministic. Since Eq. (13) then removes anchors with fewer than five local hits, low sampling probability can itself cause elimination, independently of an anchor’s realizability.

Questions for authors
  1. How does downstream co-design performance change when the topology/tree regularizers or the proximity bonus are removed separately, with the remaining training setup fixed?
  2. Are the returns in Eq. (14) normalized or temperature-scaled? How are anchors with too few sampled rollouts treated, and what prevents an empty active set?
  3. Does Table 3 report task return alone, excluding the proximity bonus? Can the final policies of all baselines be compared using the repeated evaluation protocol of Table 8, which currently reports only GeCode?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 1XGU (1/3)

Dear Reviewer 1XGU,

Thank you for your careful review and constructive feedback. We appreciate your positive assessment of the paper’s clarity and practical contributions, particularly your recognition of the shared Morphogene conditioning mechanism.

Below, we respond to your comments in detail. The training curves for the tabulated results are available as 1XGU_Fig1, 1XGU_Fig2, and 1XGU_Fig3 at the following Anonymous Link.

🖼️ Anonymous Link for Figures
👉 https://anonymous.4open.science/r/GeCode-BC8B/rebuttal/Re_1XGU.pdf

Weakness1 & Question1: Ablations of the Topology/Tree Regularizers, Proximity Bonus, and Additional Morphology Features

Thank you for raising this point. We conducted additional ablations on CHEETAH, varying one component at a time while keeping all other training settings fixed.


(1) Topology- and Tree-structured Regularization.
The two structural objectives shape the morphology space at complementary levels: Ltopo\mathcal{L}_{\mathrm{topo}} clusters morphologies with identical topologies, while Ltree\mathcal{L}_{\mathrm{tree}} encourages the latent space to preserve relative structural similarities measured by tree-edit distance.

We ablate each objective independently to assess its effect on downstream co-design performance. The results are summarized below, with the corresponding training curves available as 1XGU_Fig1 via the anonymous link.

TaskGeCodeGeCode-w/o Topo. Ltopo\mathcal{L}_{\mathrm{topo}}GeCode-w/o Tree Ltree\mathcal{L}_{\mathrm{tree}}BodyGen
CHEETAH19888.69 ± 944.6014403.66 ± 1037.1917635.99 ± 1170.2611475.06 ± 489.46

Removing Ltopo\mathcal{L}_{\mathrm{topo}} and Ltree\mathcal{L}_{\mathrm{tree}} reduces the mean return by 27.6% and 11.3%, respectively. These results indicate that both objectives contribute to downstream co-design performance in addition to shaping the latent morphology space.


(2) The Proximity Reward.
The proximity reward encourages the morphology-design policy to generate variants near its conditioning Morphogene in latent space, promoting local exploration of structurally related designs.

We remove the proximity reward and report the results in the table below. The corresponding training curves are available as 1XGU_Fig2 via the anonymous link.

TaskGeCodeGeCode-w/o Prox. λhit\lambda_{\mathrm{hit}}BodyGen
CHEETAH19888.69 ± 944.6016433.66 ± 1997.9311475.06 ± 489.46

Removing the proximity reward reduces the mean return by 17.4%, confirming the benefit of explicit locality guidance. Nevertheless, GeCode-w/o Prox. still outperforms BodyGen, in part because task returns can indirectly guide local exploration by reinforcing high-performing variants generated near each Morphogene.

(3) Additional Morphology Features.
The latent representation underlying Morphogene compactly encodes graph topology, limb geometry, and joint attributes. The observed kinematic states of the limbs and joints already convey much of this morphology information implicitly, so the encoding adds relatively little task-relevant information.

To isolate the effect of the additional morphology features, we encode the same features used to construct Morphogenes and provide them to BodyGen's controller. We denote this variant BodyGen-w/ Morpho. Features. The table below reports the results, and the corresponding training curves are available as 1XGU_Fig3 via the anonymous link.

TaskGeCodeBodyGenBodyGen-w/ Morpho. Features
CHEETAH19888.69 ± 944.6011475.06 ± 489.4612531.19 ± 370.37

Adding the morphology features increases BodyGen’s mean return by 9.2%, whereas GeCode improves on BodyGen by 73.3%. This gap suggests that the additional features account for only a small fraction of GeCode’s gain.

AuthorsOpenReview ↗

Response to Reviewer 1XGU (2/3)

Weakness2 & Question2: Eq. (14) normalized or temperature-scaled

(1) Return Normalization and Temperature Scaling.
We thank the reviewer for pointing out that the original Eq. (14) omitted return normalization and temperature scaling. As you anticipated, at each epoch we first standardize the mean returns of the top rollouts across active Morphogenes, then apply a temperature-scaled softmax (lines 131–138 of design_opt/utils/gene_utils.py):

\qquad \mathbf p = \operatorname{softmax}\\!\left(\frac{\hat{\mathbf q}}{T_p}\right),$$ where $T_p$ is the anchor-sampling temperature. Consequently, the sampling probabilities are invariant to the absolute scale and offset of the task returns, preventing the near-deterministic behavior that could result from applying softmax directly to raw returns. We have revised Eq. (14) to make both steps explicit. **(2) Anchors with Few Sampled Rollouts.** **Whether an anchor is kept depends on both local realizability and task performance.** Consider an anchor that reliably generates nearby morphologies *whose task returns are consistently low*. The anchor is locally realizable, but its designs offer limited value for co-design. Once the anchor has accumulated enough local hits, Eq. (14) uses the best returns among those hits to compute its sampling probability. If those returns remain low, the anchor is sampled less often and may eventually fall below the local-hit threshold and be removed under Eq. (13). We also enforce **a minimum sampling probability** (`min_prob` in `design_opt/conf/config.yaml`) for every active anchor to reduce the risk that insufficient sampling alone causes it to fall below the local-hit threshold. This lower bound preserves opportunities to evaluate lower-scoring anchors while allowing GeCode to allocate more rollouts to anchors that have generated high-performing local variants. **(3) Prevention of an Empty Active Set.** As stated in Lines 360–361, *“This procedure progressively eliminates Morphogenes that cannot be reliably realized by the morphology-design policy **until only one remains**.”* The elimination procedure therefore always retains one Morphogene, ensuring that the set of anchors itself cannot become empty. We also consider an edge case, ***not observed in our experiments***, in which the final Morphogene accumulates fewer than $H_{\min}$ local hits and thus falls short of the local-activity criterion. In this case, the final Morphogene remains in the anchor set and continues to condition all rollouts. With only one anchor, GeCode uses the global-best morphology as the reference for both guidance terms in Eq. (12), setting $\mathbf{z}^{\mathrm{loc}}=\mathbf{z}^{\mathrm{glob}}$. This fallback also keeps the anchor set ***nonempty*** and allows co-design to continue under global-best guidance.
AuthorsOpenReview ↗

Response to Reviewer 1XGU (3/3)

Question 3: Exclusion of the Proximity Bonus in Table 3 and Repeated Evaluation of All Baselines in Table 8

(1) Exclusion of the Proximity Bonus.
Yes. We compute all reported returns, including those in Table 3 and Fig. 4, solely from the environment’s original task reward. The proximity bonus serves only as an auxiliary training signal for the morphology-design policy. GeCode, BodyGen, and StackelbergPPO are therefore directly comparable under the same evaluation metric.

(2) Repeated Evaluation of All Baselines in Table 8.
Following your suggestion, we apply the repeated-evaluation protocol used in Table 8 to the final policies of GeCode and all baselines to enable a direct comparison of their inference-time computational costs.

TaskGeCode RewardGeCode Latency(ms)StackelbergPPO RewardStackelbergPPO Latency(ms)BodyGen RewardBodyGen Latency(ms)
cheetah17520.8070 ± 117.67400.5666 ± 0.044913006.5626 ± 81.46771.9633 ± 0.10866825.9271 ± 1122.64960.4612 ± 0.0210
terraincrosser12147.9483 ± 34.63530.6035 ± 0.00277544.0550 ± 335.13401.8874 ± 0.12454992.0994 ± 23.99110.4335 ± 0.0164
swimmer1458.0539 ± 0.66100.6083 ± 0.00091396.4335 ± 0.75731.8954 ± 0.10281314.5336 ± 0.75770.4191 ± 0.0020
glider-regular23247.6445 ± 21.89770.5431 ± 0.057712470.7871 ± 41.08481.9346 ± 0.101510995.1115 ± 40.42180.4358 ± 0.0163
glider-medium24181.9193 ± 44.27260.5242 ± 0.058813131.1225 ± 402.92171.9186 ± 0.098211938.7695 ± 75.58880.4360 ± 0.0165
glider-hard19480.9888 ± 43.06530.6104 ± 0.001314469.2543 ± 64.66301.9175 ± 0.098511726.7709 ± 259.36730.4570 ± 0.0008
walker-regular19471.2595 ± 23.27390.5246 ± 0.056315404.3003 ± 1651.92201.9015 ± 0.103911586.1626 ± 150.12270.4439 ± 0.0025
walker-medium20213.8165 ± 77.97750.6115 ± 0.004712500.5081 ± 406.65101.9312 ± 0.098213308.3472 ± 90.49830.4486 ± 0.0017
walker-hard20395.6832 ± 323.91760.5969 ± 0.001013717.2239 ± 32.05541.9219 ± 0.098812869.4025 ± 89.43520.4692 ± 0.0163
crawler17333.4610 ± 192.538615.3090 ± 6.445510856.2991 ± 100.39111.9015 ± 0.12549211.2134 ± 108.87390.4877 ± 0.0019
stepper4283.2153 ± 198.118414.8615 ± 2.56592749.3868 ± 607.09651.9770 ± 0.21911862.6032 ± 923.41040.5129 ± 0.0838
pusher4267.9213 ± 56.473310.7374 ± 3.26113642.4106 ± 23.75551.9167 ± 0.09602116.4899 ± 36.23140.4552 ± 0.0045

On the 2D tasks, where the evaluated morphologies have comparable limb counts, GeCode incurs a modest increase in FLOPs and adds only about 0.13 ms of latency per control decision on average relative to BodyGen. This suggests that Morphogene conditioning introduces little inference overhead in this setting.

On the 3D tasks, GeCode discovers higher-return morphologies with 23–29 limbs, compared with 7–14 for BodyGen. Processing these larger morphology graphs increases total FLOPs and helps explain the higher inference latency. We have expanded Table 8 to report the detailed inference-cost comparison across methods.


Dear Reviewer 1XGU,

Thank you again for taking the time to consider our response. We hope the clarifications and additional results have answered your questions 🙏. Wishing you all the best in your research ⭐️.

Reviewer_WNEAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes GeCode, which optimizes an agent’s body and controller jointly by conditioning both on a shared latent code. A handful of these codes act as anchors: designs are rewarded for staying close to their anchor, and the anchors drift toward whatever performs best.

Critical strengths
  1. Clearly specified method, extensive appendix, full hyperparameters, and anonymous code & supplementary material.
  2. Broad evaluation for this subfield: twelve tasks, five seeds, learning curves, and sample-efficiency thresholds.
  3. Ablations cover each component, and the latent-space analyses are a nice analysis.
Critical weaknesses
  1. The paper has limited novelty. Once the "gene" language is set aside, the method is a morphology autoencoder, a latent code fed to both networks (standard in universal morphology control), a FiLM-style modulation renamed AdaConcat, and a greedy update of latent prototypes that works like a simple evolution strategy.
  2. An important control is missing: condition the controller on the code of the realized morphology with no anchors and no proximity reward. Nor is a baseline controller ever trained on GeCode’s final morphologies (or vice versa). As it stands, one cannot tell whether this is a co-design result or a better-conditioned-controller result.
  3. Authors do not present a sensitivity analysis of the parameters. such as lambda_hit, t_sim, H_min, which could be insightful.
Questions for authors
  1. Can you add a variant conditioning the controller on z = E(M*) with no anchors and no proximity reward? How much of the gain survives?
  2. Can you train a baseline controller from scratch on GeCode’s final morphologies, and GeCode’s controller on theirs?
  3. What happens with lambda_hit = 0, and how sensitive are results to tau_sim and nu?
  4. Why not report results on UNIMAL or Evolution Gym?
  5. (Out of curiosity) Would the framework extend to co-designing sensors alongside limbs, and does the controller remain stable if the observation space varies across morphologies? Can authors expand on this?
6 comments
AuthorsOpenReview ↗

Response to Reviewer WNEA (1/4)

Dear Reviewer WNEA:

We sincerely thank you for your careful review and your recognition of the clarity of our methodology, the reproducibility of our study, and the breadth of our empirical evaluation.

We respond to your comments and questions below. The training curves corresponding to the tabulated results are available as WNEA_Fig1-WNEA_Fig4 at the following Anonymous Link.

🖼️ Anonymous Link for Figures
👉 https://anonymous.4open.science/r/GeCode-BC8B/rebuttal/Re_WNEA.pdf

Weakness 1: Limited Novelty

Thank you for this important question. Our main contribution is a mechanism that explicitly coordinates morphology generation and controller learning through a shared, adaptive Morphogene. Within this framework, the autoencoder defines the latent space, conditional modulation incorporates the Morphogene into both policies, and performance-based updates refine the Morphogene using rollout feedback. Together, these components implement GeCode’s novel mechanism for coordinating morphology generation and controller learning.

(1) Morphogene-guided morphology search. Following BodyGen, GeCode learns morphology construction and control within a unified reinforcement learning framework. The design policy builds each body autoregressively by selecting structural edits based on the current morphology. On its own, this stepwise exploration can concentrate on familiar structures and miss structurally distinct designs with higher returns. GeCode broadens the search by conditioning the design policy on Morphogenes that anchor distinct regions of the latent space, combining RL-based exploration within each region with global guidance across regions.

(2) Prototype-level representation and updates. In latent-space evolutionary search, each sampled code represents a single candidate morphology. In GeCode, each Morphogene serves as a prototype for a set of structurally related morphologies. For each Morphogene, the stochastic design policy generates multiple nearby variants, whose task returns are used to evaluate the anchor. GeCode combines RL-based local exploration with global search across regions. The design policy explores variants around each anchor, while the local best guides its refinement and the global best shares discoveries across anchors. Unlike standalone evolutionary search, which uses a separate candidate evaluation loop, GeCode updates its Morphogenes using rollouts already collected for RL policy training and therefore requires no additional environment interactions for anchor updates.

(3) Explicit morphology–control coordination. In BodyGen, the morphology-design and control policies are coupled primarily through shared task feedback. GeCode conditions both policies on the same Morphogene. This shared representation guides the design policy’s exploration of related morphologies and helps the controller adapt across those variants. During training, task feedback trains both policies and informs updates to the Morphogene, allowing the shared design context to adapt alongside morphology generation and control.

AuthorsOpenReview ↗

Response to Reviewer WNEA (2/4)

Weakness 2 / Question 1: Direct Conditioning Morphology Latent Code on the Controller

Thank you for raising this question. To determine whether GeCode’s improvement primarily arises from a better-conditioned controller, we construct an additional baseline, BodyGen-w/ Morphology Latent Code. Specifically, we encode the additional morphology features used to construct the Morphogene and inject the resulting representation into BodyGen’s control network using the same conditioning mechanism. Comparing this variant with standard BodyGen isolates the effect of providing additional morphology information to the controller. We summarize the results below and provide the corresponding learning curves in WNEA_Fig1, available via the anonymous link.

TaskGeCodeBodyGenBodyGen-w/ Morphology Latent Code
CHEETAH19888.69 ± 944.6011475.06 ± 489.4612531.19 ± 370.37

Adding the latent morphology representation improves BodyGen by only 9.2%, indicating that additional controller conditioning explains a limited portion of GeCode’s gain. This modest improvement is expected because the Morphogene compactly re-encodes topology, limb geometry, and joint attributes that are already largely available in BodyGen’s original morphology observations. GeCode nevertheless outperforms BodyGen-w/ Morphology Latent Code by 58.7%, demonstrating that its main advantage arises from the Morphogene’s role as a shared, searchable prototype coordinating morphology exploration with controller learning, rather than from a better-conditioned controller alone.

Weakness 2 / Question 2: GeCode's Morphology + Baseline's Controller & Baseline's Morphology + GeCode’s Controller

We thank the reviewer for suggesting this cross-evaluation. Fixing the body to a specific morphology produced by either GeCode or the BodyGen baseline reduces the original universal control problem to a single-morphology setting, in which each controller is trained specifically for the selected body.

Under this setting, the GeCode and BodyGen controllers are functionally equivalent. Their primary architectural difference is Morphogene conditioning; however, because the Morphogene is derived solely from static morphology attributes, it remains constant for a fixed body. With concatenation-based conditioning, this constant input can be absorbed into the network’s bias terms, leaving both controllers with essentially the same effective function class.

Following your suggestion, we perform a 2×22\times2 cross-evaluation, pairing the final morphologies discovered by GeCode and BodyGen with the controller architectures of both methods. For each pairing, we hold the morphology fixed and retrain the controller under the same setting.

Note that: Both controllers retain the network size used in the original co-design experiments. This relatively large capacity may make them more prone to overfitting when trained on a single fixed morphology. Nevertheless, our conclusion still holds under this shared evaluation protocol.

BodyGen ControllerGeCode Controller
BodyGen Morphology6577.43 ± 360.826743.11 ± 322.60
GeCode Morphology10690.74 ± 558.2210039.94 ± 531.15

For a given morphology, the GeCode and BodyGen controllers achieve comparable mean returns. Under both controller architectures, the morphology discovered by GeCode achieves a substantially higher mean return than the one discovered by BodyGen. These results indicate that Morphogene-driven co-design explores the morphology space more effectively and discovers high-performing body designs not identified by BodyGen. These results further highlight that GeCode’s advantages do not come from a better-conditioned controller.

AuthorsOpenReview ↗

Response to Reviewer WNEA (3/4)

Weakness 3 / Question 3: Sensitivity Analysis of the Parameters

Thank you for raising this question. Following your suggestion, we conducted additional ablation studies and sensitivity analyses for the parameters you highlighted.

(1) Proximity Reward λhit\lambda_{\mathrm{hit}}.

The proximity reward explicitly encourages the morphology-design policy to generate variants close to the conditioning Morphogene, thereby maintaining the locality and structural coherence of Morphogene-guided exploration. We remove this proximity reward by setting λhit=0\lambda_{\mathrm{hit}}=0 and holding all other settings fixed. The results are summarized below, and the corresponding learning curves are shown in WNEA_Fig2, available via the anonymous link.

TaskGeCodeGeCode (λhit=0\lambda_{\mathrm{hit}}=0)BodyGen
CHEETAH19888.69 ± 944.6016433.66 ± 1997.9311475.06 ± 489.46

Setting λhit=0\lambda_{\mathrm{hit}}=0 reduces the mean return by 17.4%, confirming the benefit of explicit locality guidance. Nevertheless, it still outperforms BodyGen by 43.2%, because locality is indirectly preserved through consistent Morphogene conditioning and similarity-based local-hit filtering, while task returns favor high-performing variants within each local region.

(2) Similarity Threshold τsim\tau_{\mathrm{sim}}.

The similarity threshold τsim\tau_{\mathrm{sim}} determines whether a generated morphology constitutes a local hit for its conditioning Morphogene, thereby defining the extent of the region anchored by that Morphogene. A lower threshold admits a broader range of morphologies but may reduce structural coherence within the region. An excessively high threshold may yield too few local hits to evaluate the anchor reliably.

We vary τsim\tau_{\mathrm{sim}} while holding all other settings fixed. The results are summarized below, and the corresponding learning curves are provided in WNEA_Fig3, available via the anonymous link.

Taskτsim=−1\tau_{\mathrm{sim}}={-1}τsim=0\tau_{\mathrm{sim}}={0}τsim=0.4\tau_{\mathrm{sim}}={0.4} (default)τsim=0.6\tau_{\mathrm{sim}}={0.6}τsim=0.9\tau_{\mathrm{sim}}={0.9}
CHEETAH15692.93 ± 1007.1519257.42 ± 838.5419888.69 ± 944.6020004.48 ± 1070.9914530.75 ± 2212.52

GeCode achieves comparable mean returns at τsim∈0,0.4,0.6\tau_{\mathrm{sim}}\in\\{0,0.4,0.6\\}, indicating limited sensitivity across these settings. Performance declines at both extremes. Setting τsim=−1\tau_{\mathrm{sim}}=-1 effectively disables similarity-based filtering, while τsim=0.9\tau_{\mathrm{sim}}=0.9 imposes a strict similarity requirement that can leave too few local hits for reliable Morphogene updates.

(3) Morphogene Update Rate η\eta.

We first clarify that the manuscript does not define a parameter denoted by ν\nu (nu). The reviewer may have intended to refer to η\eta (eta), the Morphogene update rate; under this interpretation, we report the corresponding sensitivity analysis below.

The coefficient η\eta controls the update magnitude of each Morphogene under local- and global-best guidance. We vary η\eta while keeping all other settings fixed. The results are summarized below, and the corresponding learning curves are provided in WNEA_Fig4, available via the anonymous link.

Taskη=0\eta=0η=0.05\eta=0.05η=0.1\eta=0.1 (default)η=0.3\eta=0.3η=0.7\eta=0.7
CHEETAH14788.16 ± 833.1718637.89 ± 878.2619888.69 ± 944.6020039.77 ± 829.7416922.60 ± 623.84

GeCode remains stable across the moderate range η∈[0.05,0.3]\eta\in[0.05,0.3], demonstrating limited sensitivity to the exact update rate. Performance degrades only at the extreme settings η=0\eta=0 and η=0.7\eta=0.7, where Morphogene adaptation is respectively disabled or overly aggressive.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary and Thanks to the Reviewers and Chairs

Dear Senior Area Chairs, Area Chair, and Reviewers,

We thank the reviewers for their careful reading and constructive questions, and the Senior Area Chairs and Area Chair for guiding the discussion. All three reviewers expressed a positive assessment of the paper after the rebuttal.

We summarize below the additional experiments, clarifications, and manuscript revisions we completed in response to the reviewers’ feedback.

1. Prior work and benchmark scope. We revised Section 2 to further clarify how GeCode differs from prior work (Reviewer WNEA W1 and Reviewer Hc1F W1 / Q1). We also revised the manuscript to describe the benchmark’s purpose and physical scope more precisely (Reviewer Hc1F W2). We appreciate the reviewers' patient feedback, which has significantly improved the quality of our manuscript.

2. Further clarifying the respective contributions of morphology search and controller conditioning. We added a BodyGen baseline that gives the controller an encoding of the realized morphology to isolate the contribution of this additional information (Reviewer WNEA W2 / Q1, and Reviewer 1XGU W1). We also retrained both controllers from scratch on the final morphologies discovered by each method to assess how the choice of body affects performance under either controller (Reviewer WNEA W2 / Q2).

3. Component contributions and sensitivity. We added ablations to examine the roles of structural regularization (Reviewer 1XGU W1 / Q1), the proximity reward (Reviewer WNEA W3 / Q3 and Reviewer 1XGU W1 / Q1), and local and global guidance (Reviewer Hc1F Q3). We also tested sensitivity to the similarity threshold and prototype update rate (Reviewer WNEA W3 / Q3), and further clarified how anchor sampling is implemented (Reviewer 1XGU W2 / Q2).

Once again, we thank the reviewers for their thoughtful feedback and the AC for their guidance throughout the discussion. The manuscript has improved substantially as a result. Your suggestions helped us strengthen the experimental evaluation and clarify GeCode’s contribution. We are grateful for your time and consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.