Technical Summary: COSMIC – Concurrent Optimization of Structure, Material, and Integrated Control for Robotic Systems

Problem Statement

The field of robotics aims to replicate and surpass the autonomy of natural organisms, yet current robotic systems often suffer from suboptimal performance and limited diversity compared to biological counterparts. A primary limitation is the conventional design paradigm where structure, materials, and control are developed separately or sequentially. In nature, these elements co-evolve to produce specialized, adaptive organisms. In artificial systems, this separation restricts the design space and fails to capture the collective effects arising from the tight coupling of physical dynamics and control strategies.

While recent advances in soft robotics and differentiable simulation have enabled co-design for continuum systems, the co-design of discrete robotic systems (such as truss-lattice robots, legged robots, and tensegrity structures) remains challenging. These systems present a mixed-variable design space (combining discrete topology and material selection with continuous control parameters), sharp discontinuities, and highly non-convex optimization landscapes. Existing gradient-based methods often struggle with these discrete systems, and there is a lack of comprehensive frameworks that can simultaneously optimize topology, material distribution, and control policies to reveal their individual and collective contributions to system performance.

Methodology

The authors propose COSMIC, a gradient-based co-design framework for discrete truss-lattice robots. The framework unifies the optimization of three distinct entities:

1. Topology: The connectivity of the truss network.
2. Material Distribution: The assignment of material types (Void, Skeleton, Actuator) to each truss element.
3. Control Policy: A spatiotemporal neural network controller.

1. Unified Design Representation

The framework represents the robot as a graph of nodes and edges. To enable gradient-based optimization on discrete choices, the authors employ a continuous relaxation strategy:

• Mixed-Variable Relaxation: Each edge $E_{ij}$ is associated with a continuous 3D vector $z_{ij} \in [0, 1]^3$, representing the probabilities of being in a Void, Skeleton, or Actuator state.
• Softmax Projection: A scaled softmax function projects these continuous vectors into normalized material state vectors $\tilde{z}_{ij}$, approximating one-hot encodings. A scaling factor $\beta$ controls the sharpness of this projection.
• Material Properties: A library matrix $\Psi$ maps material states to mechanical properties (stiffness $k$, linear density $\rho$, and maximum actuation strain $a_{max}$). The effective properties of each truss are interpolated based on $\tilde{z}_{ij}$.
• Control Network: A Multi-Layer Perceptron (MLP) generates control signals. Inputs include Central Pattern Generator (CPG) signals, nodal positional offsets, velocities, and a goal vector. The raw MLP outputs are modulated by the material state vector, ensuring only trusses in the Actuator state receive target strain commands.

2. Differentiable Dynamic Simulation

The framework utilizes a GPU-parallel differentiable simulator (implemented in Taichi) to evaluate system dynamics.

• Physics Model: Truss forces are governed by a generalized Hooke's law, where internal forces depend on the difference between current length and a target length determined by the controller.
• Contact Handling: A high-fidelity Time of Impact (TOI) scheme is used to handle ground contacts, preventing interpenetration and reducing gradient discontinuities. This allows for the backpropagation of gradients through collision events, including frictional damping.
• Differentiability: The entire forward simulation (integration, contact resolution, and control generation) forms a computational graph, enabling automatic differentiation to compute gradients of the performance objective with respect to both structural variables and neural network weights.

3. Constrained Optimization Strategy

The co-design problem is formulated as a constrained optimization to maximize forward locomotion distance.

• Objective: Minimize the negative displacement of the payload head node.
• Constraints:
  • Volume Fractions: Bounds on total solid volume and actuator volume.
  • Binarization: To drive continuous variables toward discrete states, the framework employs a global binarization penalty (L2 norm) and a pairwise orthogonal penalty to minimize overlap between material states.
• Optimization Algorithm:
  • Augmented Lagrangian (AL): Used to handle binarization constraints adaptively, allowing temporary violations to explore the design space before enforcing strict constraints.
  • Multi-Stage Process:
    1. Co-design Phase: Alternating updates of the MLP weights (via Adam) and material design vectors (via Method of Moving Asymptotes, MMA).
    2. Stability Pass: An auxiliary evaluation using a highly sharpened projection ($\beta_{stab}$) and a Straight-Through Estimator (STE) to assess structural integrity of the emerging discrete geometry.
    3. Iterative Attraction: A phase where design vectors are progressively "nudged" toward their dominant discrete state.
    4. Fine-tuning: A final stage where the topology and material layout are frozen (hard discretized), and only the controller is fine-tuned.

Key Results

The framework was validated through extensive case studies on a 6x6 nodal truss lattice robot.

1. Performance vs. Baselines

• Co-design vs. Sequential Design: A baseline robot with a fixed, heuristic-based structure (three-legged topology with pre-defined actuator placement) achieved a displacement of 1.14 m. The co-designed robot (Design 6b) achieved 1.90 m, a 66% improvement.
• Initialization Robustness: The framework successfully discovered viable locomotion strategies even from a "no-heuristic" initialization (uniform random material ratios), achieving 0.98 m. This demonstrates the ability to navigate the non-convex design space without relying on specific initial guesses, though performance was sensitive to initialization.

2. Ablation Studies (Individual and Collective Contributions)

Combinatorial ablation studies (disabling topology, material, or control optimization) revealed:

• Material Layout: Had the largest individual impact, increasing performance by 781% compared to fixed layouts.
• Topology: Provided a 116% improvement on its own but acted as a powerful multiplier for material design.
• Control: Showed a small individual contribution and could even degrade performance if the underlying structure was suboptimal. However, when co-evolved with materials and topology, it contributed to an 87% increase in distance compared to the case without control optimization.
• Synergy: The joint optimization of all three entities yielded results significantly superior to the sum of optimizing them in isolation. The study observed "convergent evolution," where different initializations led to similar structural motifs (e.g., specific leg divisions) when optimized for the same task.

3. Generalizability

The framework adapted to varying tasks and environments:

• Inclined Terrain: The robot evolved a slimmer structure with a lower center of mass and an arching gait to climb a 15-degree incline.
• Reduced Design Space: On a smaller 5x5 grid, the framework rediscovered leg-division strategies and achieved comparable performance to larger designs.
• Friction Variations: Under Coulomb friction (low friction), the robot adopted a strategy of keeping the payload grounded and using the front leg as a friction pad, a behavior not seen in infinite-friction scenarios.

Significance and Claims

The paper claims to provide a computational foundation for the autonomous co-design of robotic systems. Its primary significance lies in:

1. Unified Framework: It successfully bridges the gap between discrete topology optimization and continuous control learning, handling the mixed-variable nature of discrete robotic systems through differentiable simulation and continuous relaxation.
2. Performance Gains: It demonstrates that co-design consistently outperforms sequential design strategies, uncovering diverse and high-performance locomotion strategies (leaping, sweeping, hopping) that are inaccessible to separated design pipelines.
3. Design Insights: It offers a quantitative understanding of how structure, material, and control interact. The study highlights that while material layout is the most critical individual factor, the synergistic coupling of all three entities is essential for achieving optimal performance.
4. Flexibility: The framework is shown to be adaptable to different functional requirements, physical environments, and design constraints, suggesting its potential for designing complex autonomous behaviors in reconfigurable and bio-hybrid robots.

The authors acknowledge limitations, including sensitivity to initialization (suggesting a need for hybrid global-local search methods) and the use of simplified physics models, but position the work as a step toward more robust, autonomous robotic design tools.