Technical Summary: Generation of Diverse and Functional Robot Designs using Superquadrics Parametrisation and Quality-Diversity

Problem Statement

The automatic generation of robots involves the simultaneous optimization of a 3D physical morphology and its controller. While Evolutionary Algorithms (EAs) are well-suited for this black-box optimization problem, they frequently suffer from premature convergence to a small set of sub-optimal or simplistic designs. This issue is exacerbated by the high dimensionality and complexity of the search space, which limits the number of designs that can be generated and tested within feasible timeframes. Furthermore, standard EAs often lack mechanisms to explicitly maintain diversity, which is crucial for downstream tasks where specific designs must be selected and fine-tuned. Additionally, common generative representations like Compositional Pattern Producing Networks (CPPNs) can introduce biases toward specific design types (e.g., symmetrical "spider" shapes) and are difficult to tune due to complex hyperparameters.

Methodology

The authors propose a novel framework, ME2HKS, which combines a specific geometric representation with a Quality-Diversity (QD) algorithm and a self-organizing control mechanism.

1. Representation: Superquadrics (SQ) vs. CPPN

The study compares two encoding strategies for the robot chassis:

• Dual CPPN: Uses two Compositional Pattern Producing Networks. One network generates the chassis shape based on spatial coordinates, and a second network determines the placement and type of components (wheels, sensors, limbs, castors) on the chassis surface.
• SQ-CPPN: Replaces the chassis-generating CPPN with a Superquadrics (SQ) representation. SQs are compact mathematical representations of 3D shapes defined by only 8 parameters (plus symmetry flags), allowing for the generation of diverse geometries (ellipsoids, cubes, etc.) with high interpretability. A single CPPN is retained to place components on the SQ-generated chassis.
  • The SQ representation divides the chassis into four quadrants, each defined by a superquadric equation, allowing for complex, non-uniform shapes while maintaining a small parameter space.

2. Algorithm: Morpho-Evolution with Homeokinesis (MEHK) and MAP-Elites

The core evolutionary process is an adaptation of the Morpho-Evolution with Homeokinesis (MEHK) algorithm:

• Outer Loop (Morpho-Evolution): Evolves the robot design. The authors compare a standard asynchronous EA (AME) against a MAP-Elites variant (AME2). MAP-Elites maintains a multi-dimensional archive (grid) where cells are defined by morphological features (number of wheels, joints, sensors, and chassis dimensions). Solutions are stored based on their fitness, ensuring that diverse morphologies are preserved even if they are not the absolute fittest.
• Inner Loop (Homeokinetic Control - HKC): Instead of using computationally expensive reinforcement learning to train controllers, the system employs a Homeokinetic Controller (HKC). This self-organizing algorithm balances the predictability and sensitivity of the robot-environment interaction. It allows any robot design to quickly generate coherent, exploratory behaviors (e.g., moving in lines or circles) without a specific goal, serving as a fast proxy for functional mobility.

3. Experimental Setup

Experiments were conducted in two environments: a Flat Arena and a Cracks Arena (simulating uneven terrain). The study evaluated four variants:

1. MEHKC: Standard MEHK with Dual CPPN.
2. MEHKS: Standard MEHK with SQ-CPPN.
3. ME2HKC: MAP-Elites MEHK with Dual CPPN.
4. ME2HKS: MAP-Elites MEHK with SQ-CPPN.

Each variant ran 20 times with a budget of 10,000 design evaluations per run. Metrics included Archive Coverage (number of filled cells in the 6D grid), Quality-Diversity (QD) score (sum of fitness values in the archive), and Maximum Fitness.

Key Contributions

1. Novel Representation in Morpho-Evolution: This is the first application of Superquadrics within a nested morpho-evolution algorithm. The authors demonstrate that SQs provide a more interpretable and computationally efficient representation for chassis design compared to CPPNs, reducing the search space while maintaining geometric expressiveness.
2. Integration of QD with MEHK: The authors adapt the MEHK algorithm to function as a Quality-Diversity algorithm by integrating MAP-Elites. This explicitly enforces morphological diversity by mapping the population onto a 6-dimensional archive of design features.
3. Efficient Evaluation via Homeokinesis: The study validates the use of Homeokinetic Control as a rapid, goal-agnostic method to evaluate the functional potential of diverse robot morphologies, enabling the generation of thousands of designs in a fraction of the time required by deep reinforcement learning methods.

Results

The ME2HKS variant (SQ-CPPN + MAP-Elites) consistently outperformed all other configurations:

• Diversity: ME2HKS achieved the highest Archive Coverage and QD scores in both environments. Specifically, using SQ-CPPN generated four times more unique designs in the archive compared to Dual CPPN. Adding MAP-Elites to the SQ-CPPN approach doubled this number again.
• Quality: While the maximum exploration scores (fitness) were statistically similar across all variants (approx. 50% of the flat arena and 14-19% of the cracks arena visited), the QD score increased proportionally with archive coverage. This indicates that ME2HKS maintained high functional quality while exploring a significantly larger portion of the design space.
• Morphological Bias: The results showed that the algorithm correctly adapted to the environment: robots in the flat arena predominantly evolved wheels, while those in the cracks arena evolved limbs to prevent getting stuck. The SQ-CPPN representation allowed for a broader distribution of high-performing designs with varying component configurations (e.g., higher numbers of limbs and sensors) compared to the Dual CPPN.

Significance and Claims

The paper claims that the combination of a compact, interpretable geometric representation (Superquadrics) and an explicit diversity mechanism (MAP-Elites) significantly enhances the generation of functional robot designs.

• Interpretability and Efficiency: The SQ representation allows for easier tuning and navigation of the design space compared to the complex hyperparameter tuning required by CPPNs.
• Scalability: By combining SQs with the fast HKC controller and MAP-Elites, the system can discover nearly 4,000 unique, functional designs in a single run (using 64 CPUs), overcoming the data-efficiency limitations typically associated with MAP-Elites.
• Downstream Utility: The generated archive provides a diverse set of functional morphologies, facilitating the selection and fine-tuning of robots for specific downstream tasks.

The authors conclude that while MAP-Elites is often computationally expensive, the efficiency of the homeokinetic controller makes it a viable strategy for exploring complex morphological spaces, provided a compact representation like Superquadrics is used to guide the search.