- 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:
- Hu et al. (2023) CoRL. https://proceedings.mlr.press/v205/hu23c.html
- Li et al. (2025) ICLR. https://openreview.net/forum?id=awvJBtB2op
- Yu et al. (2026) PNAS. https://www.pnas.org/doi/10.1073/pnas.2519129123
- Wang et al. (2026) ICML. https://openreview.net/forum?id=dTIxiQ0jhu
- 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
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.
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 , while our evolved CHEETAH agent reaches approximately (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 – and a volume-weighted density of , which is close to that of solid ABS engineering plastic (approximately ).
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.
| Setting | Max gear | Limb density range () | Volume-weighted density () |
|---|---|---|---|
| Original benchmark | 400 | 952.8–986.2 | 982.0 |
| 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 gear | Max. joint torque () | Max. single-joint power () | Transform2Act | BodyGen | StackelbergPPO | GeCode | |
|---|---|---|---|---|---|---|---|
| Original benchmark | 400 | 400.00 | 29.29 | 9542.1 ± 635.0 | 11475.1 ± 489.5 | 13246.3 ± 299.9 | 19888.7 ± 944.6 |
| Restricted gear | 100 | 100.00 | 1.66 | 2311.16 ± 246.85 | 3350.65 ± 562.44 | 4255.77 ± 788.26 | 7956.38 ± 749.69 |
| Restricted gear | 50 | 50.00 | 0.70 | 1042.07 ± 39.19 | 1013.70 ± 109.91 | 1093.37 ± 30.34 | 3316.09 ± 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) | GeCode | StackelbergPPO | BodyGen | Transform2Act |
|---|---|---|---|---|
| Training time (hours) | 22.952 ± 0.760 | 23.421 ± 0.514 | 20.260 ± 0.467 | 12.460 ± 0.521 |
| Final return | 15885.39 ± 2400.07 | 9140.41 ± 1354.80 | 8242.09 ± 1320.33 | 5194.98 ± 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.
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
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).