Co-design of Neural and Muscle Network based on Embodied Perceptron Representation

AI summary

Overview

Research area: Robotics — embodied intelligence, musculoskeletal robot co-design, and biologically inspired control. Published in the Proceedings of the IEEE/SICE International Symposium on System Integration (SII 2026), DOI: 10.1109/SII64115.2026.11404727 (arXiv:2608.16555v1 [cs.RO]).

Technical level: Intermediate. The paper uses neural-network notation and matrix representations that assume some familiarity with deep learning and robot dynamics, but the core idea is explained through a direct body-to-network analogy.

One-sentence scope: The paper proposes the Embodied Perceptron, a representation that writes a physical musculoskeletal body as a neural network, and uses it to jointly optimize a walking robot's muscle configuration and its control policy.

What This Paper Is About

Robot bodies are usually built as simple serial chains of motors and rigid links, because simple bodies are easy to design — but that simplicity can limit how well a robot adapts to its environment. Embodied robotics has shown that a well-designed body can take over part of the control and computation burden through physical body–environment interaction, yet such designs still depend heavily on expert intuition, and there is no unified theory that places the controller and the body on equal footing. This paper supplies such a framework — the Embodied Perceptron — and demonstrates it by co-optimizing muscle layout and control in a musculoskeletal walking robot.

Key Contributions

1. A unified representation. The Embodied Perceptron formally maps a physical body system onto the structure of a feedforward neural network: mechanical parameters (such as muscle configuration) correspond to weights, and physical nonlinearities (such as nonlinear muscle elasticity) correspond to activation functions. The paper shows the analogy is formal, not merely metaphorical.
2. An explicit account of control substitution. Because the body can be written in neural-network form, the framework allows the authors to state explicitly and theoretically that a body can substitute for part of the neural control.
3. Joint optimization of body and controller. The authors co-optimize three things at once in a simplified musculoskeletal robot implemented in MuJoCo: the neural control policy, the muscle configuration (Jacobian matrix R), and muscle properties (stiffness matrix K), using CMA-ES.
4. A demonstration that minimal controllers suffice. The optimized body allows stable walking with a single-neuron controller, whereas conventional two-layer feedforward networks needed many more neurons (often exceeding 16 in total) to reach comparable locomotion performance.

Main Findings

• Body-plus-controller learns walking with extremely small networks. Across hidden-layer sizes from 1 to 32 neurons over 1000 training iterations (averaged over ten runs, with error bars for variability), reward generally increased and stabilized regardless of network size. Networks with as few as one or two neurons in the hidden layer achieved stable learning and long travel distance, with reward levels comparable to larger networks.
• More neurons did not help at 1000 iterations. Examining travel distance at 1000 iterations, increasing the number of neurons (greater than 16) tended to slightly reduce learning efficiency.
• Co-optimized body parameters are reported. With a single neuron in the controller, the optimized muscle configuration matrix R (10 × 3) and stiffness matrix K = diag([323.13, 617.18, 296.07, 94.18, 1391.27, 4391.81, 4617.86, 3807.32, 5724.71, 7161.96]) are given. Rows with negative entries indicate flexor action, positive entries extensor action; K values are positive stiffness coefficients, with darker colors in the visualization indicating higher elasticity.
• Body-aware design beats plain networks. Compared with conventional two-layer feedforward networks, which required significantly more neurons (often exceeding 16 in total) for comparable performance and which showed poor learning efficiency and limited convergence when small, the proposed architecture achieved stable locomotion learning with as few as 1–4 neurons.
• A scaling identity keeps parameters physically plausible. Because no upper bound is placed on R entries, the paper introduces an equivalence transformation (R̂ = aR, K̂ = a⁻³ᐟ²K, b̂ = ab, f̂_active = a⁻³ᐟ²f_active). For instance, if the largest optimized R element r_max exceeds 0.1, choosing a = 0.1/r_max rescales pulley radii so the maximum becomes 0.1.
• The physical body is a computational resource, not a passive structure. The paper argues the results point toward memory-efficient and data-efficient learning through embodied co-design, and draws a parallel with biological systems where simple neural circuits such as spinal reflexes generate complex gait patterns, suggesting inter-joint coordination can emerge from body–environment interaction.

Methodology in Plain English

The authors start from an observation about muscles. A muscle's length is approximately a linear function of joint angles, and the torque that muscles place on joints is that same relationship transposed and multiplied by muscle tension. Tension itself splits into a passive part — produced by elastic tissue, and behaving nonlinearly — and an active part produced by the controller.

They then write these relationships out as a sequence of network layers: a linear layer turning joint angles into muscle extensions (with the Jacobian matrix R acting as weights and a bias term from initial muscle length), a nonlinear activation layer applying a squared-ReLU (ReLU²) to represent passive force with slack, and a final linear layer mapping tension back to joint torque via R transpose. The active muscle force is supplied by a separate neural network that plays the role of the control policy. The robot's onboard computer therefore only needs to run the small controller; the rest is the physical mechanism.

For the experiment, they build a front-wheeled, rear-legged hybrid robot in MuJoCo, with ten muscles per leg, giving an R matrix of size 10 × 3 and a diagonal K matrix of size 10 × 10. Joint angles are randomly sampled at initialization (image posture as zero, radians, extension positive, flexion negative), and the random seed is fixed for reproducibility. All experiments ran on an Ubuntu 22.04 workstation with an Intel Core i9-14900K processor.

Optimization uses CMA-ES rather than gradient descent, because CMA-ES does not require differentiable objectives and the authors consider it more robust to noisy or discontinuous reward landscapes, which are common when morphology and control are co-optimized. The reward combines total walking distance with three penalties: an energy-consumption penalty on absolute joint power summed over six joints, penalties on negative elements of R and K, and an L1 penalty on positive R elements to encourage sparsity. The walking task's success criterion is exceeding 15 m within 5000 simulation steps.

Why This Matters

Impact on research. The paper reframes the classic question of "how much can the body do instead of the brain" as an explicit representational question. By writing the body as a network, it gives a formal language in which the body's contribution to computation can be stated, and it makes controller-and-body co-design a single optimization problem rather than two separate expert-driven ones. It also connects to earlier threads in the literature — physical reservoir computing, passive dynamic walkers, reinforcement learning for tendon-driven underactuated robots, and optimizing multi-articular muscle arrangement from human motion data — under one framing.

Real-world applications:

• Design of musculoskeletal and tendon-driven robots, where muscle routing and stiffness are currently chosen largely by intuition.
• Bio-inspired legged and hybrid wheel-leg locomotion platforms that must operate with limited onboard computation.
• Embedded and resource-constrained controllers, since the framework shows a walking policy can be realized with a single neuron when the body is designed to carry part of the load.
• Systematic exploration of morphology for prostheses and assistive mechanisms, where the trade-off between passive mechanical behavior and active control is central.

Industry relevance. Anywhere hardware and control are co-developed under cost, weight, or compute constraints — legged logistics robots, agricultural and inspection platforms, wearable assistive devices — the paper suggests that investing in the mechanical design can directly reduce the size of the learned controller, which affects memory footprint, training data needs, and inference cost.

Future Directions

• Extending beyond walking. The paper notes the framework is general and can be extended to goal-directed motions by redefining the reward and the control objectives; the current study focuses on a specific locomotion task.
• Scaling to more complex robots. Future work aims to scale the approach to more complex robots, integrating diverse physical constraints and environmental factors.
• Replacing soft constraints with feasible ones. The penalty-based body parameter optimization currently uses soft constraints and can occasionally yield infeasible configurations. Increasing optimization iterations can mitigate this, but hyperparameter selection remains a key challenge.
• Understanding when neural control is still needed. The finding that coordination may emerge from body–environment interaction without complex neural control raises the open question of exactly which control functions the body can absorb and which it cannot.

Target Audience

Robotics researchers working on embodied intelligence, morphological computation, and co-design of morphology and control; engineers building musculoskeletal, tendon-driven, or bio-inspired walking robots; and machine learning researchers interested in how physical structure can substitute for network capacity. Readers with a background in neural networks and robot dynamics will get the most from the theory section, while the results and discussion are accessible to a broader robotics audience.

Authors’ abstract

Recent advances in AI technologies have enabled the advanced design of complex control policies. In contrast, focusing on the body, many robots still employ simple bodies that can limit adaptability to environments. Studies in embodied robotics have shown that well-designed bodies can partially replace the role of control and computation with physical body-environment interactions, yet such designs still depend heavily on expert intuition. There is a need for a systematic theoretical framework for body design, as well as a method for joint optimization of the body and controller. To address this, we introduce the Embodied Perceptron, a theoretical framework that unifies neural networks and physical body systems. In this view, the body itself acts as a perceptron: mechanical parameters correspond to weights, and physical nonlinearities play the role of activation functions. By representing physical constraints as weights and nonlinear properties as activation functions, a physical body can be modeled in neural-network form. The system representation enables us to explicitly and theoretically explain that the body can substitute for part of the neural control. As an application, we co-optimize control policy and muscle configuration in a musculoskeletal robot and show that the resulting embodied intelligence can provide inherent stability, improve learning efficiency, and drastically reduce model size-even with a single-neuron controller. The results bridge the informational and physical worlds and provide a pathway toward understanding and systematic design of embodied AI systems.