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

Fu Feng; Ruixiao Shi; Yucheng Xie; Jing Wang; Xin Geng

cs.LGSep 25, 2026

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

Authors: Fu Feng, Ruixiao Shi, Yucheng Xie, Jing Wang, Xin Geng

Organizations: School of Computer Science and Engineering, Southeast University, Nanjing, China · Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China

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.

Figures & tables

Figure 1: (a) Traditional morphology–control co-design employs separate morphology-design and control networks coupled primarily through a shared task objective, resulting in weak body–brain coordination. (b) Inspired by biological genes, GeCode introduces Morphogene as a high-level latent blueprint that bridges body and brain. By propagating Morphogene variations to both morphology design and controller adaptation, GeCode drives their explicit coordination at the representation level while reformulating morphology search as exploration within a continuous latent space.

Figure 2: Overview of Morphogene space construction and conditioning. (a) An RAE encodes randomly generated morphologies into morphology codes, forming a continuous latent space in which Morphogenes serve as searchable anchors. (b) AdaConcat adapts an agent-level Morphogene into limb-specific conditions for the morphology-design and control networks.

Figure 3: Overview of GeCode. GeCode reformulates morphology–control co-design as exploration over the compact Morphogene space, using a small set of Morphogenes as searchable design anchors. At each co-design epoch, sampled anchors jointly condition morphology generation and controller adaptation, while stochastic rollouts guided by a proximity objective explore their local neighborhoods. Rollout performance subsequently drives Morphogene updates and adaptive resampling, propagating high-performing design information across the set and enabling global latent-space exploration.

Figure 4: Performance comparison of GeCode with state-of-the-art baselines, including StackelbergPPO ( Dai et al., 2026 ) , BodyGen ( Lu et al., 2025 ) , and Transform2Act ( Yuan et al., 2022 ) , and its ablations, GeCode-w/o Gene2Body and GeCode-w/o Gene2Brain, across twelve co-design environments. Shaded regions indicate the standard error over five random seeds.

	Crawler	Stepper	Pusher	Cheetah	TerrainCrosser	Swimmer
w/ Direct Concat	14049.3 $\pm$ 533.4	3296.0 $\pm$ 560.1	4289.5 $\pm$ 989.3	18734.4 $\pm$ 523.1	7826.2 $\pm$ 798.0	1353.5 $\pm$ 27.4
w/ AdaConcat	16143.3 $\pm$ 583.0	3688.6 $\pm$ 271.0	6250.2 $\pm$ 757.9	19888.7 $\pm$ 944.6	9670.6 $\pm$ 1176.4	1422.5 $\pm$ 29.5
	Glider-Regular	Glider-Medium	Glider-Hard	Walker-Regular	Walker-Medium	Walker-Hard
w/ Direct Concat	20998.1 $\pm$ 1002.0	20333.9 $\pm$ 884.1	18031.0 $\pm$ 1571.1	19197.0 $\pm$ 2119.3	19949.6 $\pm$ 856.5	20647.6 $\pm$ 851.3
w/ AdaConcat	23317.5 $\pm$ 627.9	22376.4 $\pm$ 809.9	19972.8 $\pm$ 332.1	24306.1 $\pm$ 1293.0	21973.3 $\pm$ 918.9	21614.6 $\pm$ 521.8

Table 1: Performance comparison of different Morphogene injection mechanisms. Results show mean episode rewards and standard errors over five random seeds.

Figure 5: (a) Representative morphologies designed by GeCode. (b) Visualization of limb-specific conditioning produced by AdaConcat. (c) Ablations on Morphogene updates and set size $N_{G}$ .

Figure 6: (a) Ablation study of structural constraints (i.e., $\mathcal{L}_{\mathrm{topo}}$ and $\mathcal{L}_{\mathrm{tree}}$ ) in the latent space. We report topology-aware 1-NN consistency and the Spearman rank correlation between latent and tree-edit distances, denoted “1-NN $\uparrow$ ” and “Spearman $\uparrow$ ”, respectively. (b) Morphogene update trajectories over 1,000 co-design epochs. and denote the initial and final positions, respectively, while numerical labels indicate elimination epochs. See Figure 14 in App. E.4 for additional trajectories.

Figure 7: Computational efficiency of GeCode. (a) Performance comparison under a wall-clock time budget. (b) Detailed training costs, together with the resources required for GeCode to match each baseline’s best performance. Further results are provided in Table 7 and Table 8 in App. E.3 .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Overview of Task-Specific Initial Morphologies. Initial Type 1 is a three-dimensional morphology comprising four limbs. Initial Type 2 comprises two limbs connected by one joint in the $x$ – $y$ plane. Initial Type 3 comprises two limbs connected by one joint in the $x$ – $z$ plane. Initial Type 4 comprises three limbs connected by two joints in the $x$ – $z$ plane. These morphologies serve as the starting points for sequential morphology design.

Figure 9: Overview of the Benchmark Environments. Crawler , Stepper , and Pusher are three-dimensional (3D) tasks, whereas the remaining environments are two-dimensional (2D), operating in either the $x$ – $z$ or $x$ – $y$ plane. The benchmarks span diverse morphology depths, symmetries, limb configurations, terrains, and task objectives, enabling a comprehensive evaluation of morphology–control co-design across locomotion and manipulation settings.

Panel of Table 2 — Table 2: Hyperparameters of GeCode adopted in all experiments.

Robot Morphology-Control Co-Design

Robot morphology-control co-design studies how to jointly shape a robot’s body and controller to improve task performance.

See all 59 papers

Efficient Morphology-Control Co-Design via Stackelberg Proximal Policy Optimization
Mar 16, 2026Yanning Dai, Yuhui Wang, Dylan R. Ashley +1Robot Morphology-Control Co-Design Bilevel Optimization
Center of Excellence for Generative AI, King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia. · Dalle Molle Institute for Artificial Intelligence Research (IDSIA), Lugano, Switzerland. · Università della Svizzera italiana (USI), Lugano, Switzerland. +1
Evolving Embodied Intelligence: Graph Neural Network--Driven Co-Design of Morphology and Control in Soft Robotics
Mar 20, 2026Jianqiang Wang, Shuaiqun Pan, Alvaro Serra-Gomez +2Robot Morphology-Control Co-Design Robot Policy Adaptation
LIACS, Leiden University · School of Mechanical Engineering, Xi’an Jiaotong University · Department of Computer Science, Loughborough University
Shaping the Evolutionary Dynamics of Robot Morphology via Adaptive Control Learning
Aug 24, 2026Junru Song, Yang Yang, Yaqing Xu +5Robot Morphology-Control Co-Design Evolutionary Robotics
Shanghai Jiao Tong University, Shanghai 200240, China · Intelligent Game and Decision Laboratory, Beijing 100048, China · Renmin University of China, Beijing 100872, China +1

Mar 16, 2026J/K move · Enter open · S save

Efficient Morphology-Control Co-Design via Stackelberg Proximal Policy Optimization

Yanning Dai, Yuhui Wang, Dylan R. Ashley, Jürgen Schmidhuber

Center of Excellence for Generative AI, King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia. · Dalle Molle Institute for Artificial Intelligence Research (IDSIA), Lugano, Switzerland. · Università della Svizzera italiana (USI), Lugano, Switzerland.+1

Robot Morphology-Control Co-Design Bilevel Optimization Proximal Policy Optimization

Open paper AskPDF arXiv

Morphology-control co-design concerns the coupled optimization of an agent's body structure and control policy. This problem exhibits a bi-level structure, where the control dynamically adapts to the morphology to maximize performance. Existing methods typically neglect the control's adaptation dynamics by adopting a single-level formulation that treats the control policy as fixed when optimizing morphology. This can lead to inefficient optimization, as morphology updates may be misaligned with control adaptation. In this paper, we revisit the co-design problem from a game-theoretic perspective, modeling the intrinsic coupling between morphology and control as a novel variant of a Stackelberg game. We propose Stackelberg Proximal Policy Optimization (Stackelberg PPO), which explicitly incorporates the control's adaptation dynamics into morphology optimization. By modeling this intrinsic coupling, our method aligns morphology updates with control adaptation, thereby stabilizing training and improving learning efficiency. Experiments across diverse co-design tasks demonstrate that Stackelberg PPO outperforms standard PPO in both stability and final performance, opening the way for dramatically more efficient robotics designs.

Abstract

Robot Morphology-Control Co-Design

Explore similar work

Efficient Morphology-Control Co-Design via Stackelberg Proximal Policy Optimization

Evolving Embodied Intelligence: Graph Neural Network--Driven Co-Design of Morphology and Control in Soft Robotics

Shaping the Evolutionary Dynamics of Robot Morphology via Adaptive Control Learning