💻 computer science

Graph-Based Design of Soft Grippers with Multi-Objective Quality-Diversity Optimisation

This paper proposes a graph-based design space coupled with a multi-objective, diversity-driven genetic optimization framework to generate soft grippers that exhibit emergent generalization and improved robustness across diverse, unseen grasping scenarios.

Original authors: Andre Farinha, Ge Shi, Harry Bowman, Brendan Tidd, David Howard, Josh Pinskier

Published 2026-09-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andre Farinha, Ge Shi, Harry Bowman, Brendan Tidd, David Howard, Josh Pinskier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where robots can pick up a ripe strawberry without bruising it, then immediately switch to lifting a heavy, irregularly shaped rock, all without needing a new set of instructions or a perfect camera view. This is the holy grail of robotics: a machine that can handle the messy, unpredictable variety of the real world. For decades, engineers have tried to solve this by building robots that think harder, using complex software to calculate every movement. But there is another path, one that relies on the physical body of the robot itself. Soft robots, made from flexible materials like rubber or silicone, naturally bend and squish to fit around objects. They possess a kind of mechanical intelligence, where the shape of the robot helps it do the job. However, designing these soft hands is incredibly difficult. Because they can bend in so many ways, there are nearly infinite ways to build them, and finding the one that works best for everything is like searching for a needle in a haystack that keeps changing shape.

A team of researchers at CSIRO Robotics in Australia has taken a fresh approach to this problem. Instead of trying to design a single perfect gripper for a specific task, they created a system that designs a whole family of diverse soft grippers at once. They treated the design of these robots like a puzzle made of connected points and lines, a method that allows them to explore a vast number of shapes without getting stuck on just one solution. By testing these designs against many different scenarios—grabbing small hard objects, large soft ones, and everything in between—they discovered something surprising. When they forced their computer to find solutions that worked well across all these different situations, the resulting robots didn't just become good at one thing; they became surprisingly good at handling objects they had never seen before.

The researchers started by building a digital language to describe these soft robots. Rather than drawing a solid shape, they represented the robot as a network of dots and lines. The dots represent points in space, and the lines represent the material connecting them. This network is flexible; the computer can add or remove dots, move them around, or change how thick the connecting lines are. This approach is powerful because it allows the computer to invent shapes that a human designer might never think of, creating complex curves and structures that are still simple enough for the computer to test quickly. To make sure these digital designs were realistic, the team used advanced physics simulations that account for how soft materials stretch, twist, and bounce back. They even simulated how the robot touches an object, ensuring the virtual rubber didn't pass through the virtual apple or rock.

The real innovation, however, was in how they searched for the best designs. Usually, when computers try to solve a problem, they focus on finding the single best answer for a specific task. If you ask a computer to design a gripper for a tennis ball, it will create a shape perfect for a tennis ball but useless for anything else. This is called overfitting, and it is a major hurdle in robotics. The CSIRO team decided to fight this by using a strategy that values variety. They asked their computer to find designs that were not only good at grabbing, but also different from one another. They set up a competition where the computer had to juggle four different goals at once: grabbing a small hard object, a large hard object, a small soft object, and a large soft object. The computer was told to keep a wide range of solutions, ensuring that no single type of gripper dominated the results.

As the computer ran thousands of simulations, evolving these designs over hundreds of generations, a clear pattern emerged. The designs that were forced to be diverse did not just become a messy mix of mediocre solutions. Instead, they developed a robustness that single-task designs lacked. When the researchers tested these diverse designs on objects they had never shown the computer during the training phase, the results were striking. The grippers designed for a single task failed miserably when faced with a new shape. But the grippers that had been optimized for variety handled these new challenges with ease. They could adapt to a concave column or a coral-like structure that looked nothing like the training objects. It turned out that by teaching the computer to be flexible in its thinking, the physical robots learned to be flexible in their bodies.

To prove this was not just a computer artifact, the team built two of these optimized grippers out of real rubber. They mounted them on a motor and tested them by pulling on various objects, including a sphere, a pin, and a delicate coral-like structure. The physical robots behaved almost exactly as the simulations predicted. They used their softness to conform to the shapes of the objects, generating enough force to hold them without crushing them. This confirmed that the digital designs had captured real, physical principles of how soft materials interact with the world. The experiment showed that the "mechanical intelligence" of the robot was doing the heavy lifting, allowing it to generalize its skills to new situations without needing to be reprogrammed.

The study suggests that the key to building truly adaptable robots might not be in making them smarter, but in making their design process more diverse. By refusing to settle for a single perfect solution and instead embracing a wide range of possibilities, engineers can create machines that are naturally robust against the unexpected. This approach offers a promising path forward for robots that need to work in unpredictable environments, from harvesting fruit in a field to assisting in a laboratory. It shows that sometimes, the best way to prepare for the unknown is to practice with a little bit of everything.

Technical Summary: Graph-Based Design of Soft Grippers with Multi-Objective Quality-Diversity Optimisation

Problem Statement

Effective manipulation across diverse objects is critical for applications ranging from agricultural harvesting to laboratory automation. While soft robotics offers inherent compliance suitable for this challenge, designing grippers that generalize across tasks remains difficult. The design space of continuum mechanics is vast, and standard optimization approaches often lead to overfitting on specific scenarios. Furthermore, there is a fundamental trade-off in computational design: restricting design spaces ensures compatibility with high-fidelity simulation but limits expressiveness, while exploring richer spaces with simplified physics neglects the non-linear phenomena (large deformations, material non-linearities, instabilities, and contact interactions) that govern soft gripper performance. Additionally, conventional single-objective optimization frequently yields narrow, brittle solutions that exploit specific conditions rather than covering diverse behavioral spaces.

Methodology

1. Graph-Based Design Representation

The authors propose a graph-based design space for planar soft structures and compliant mechanisms that balances expressiveness with tractability for finite element simulation. The representation consists of three layers:

Genotype: An information-dense internal representation composed of $n$ nodes, where each node $i$ is defined by Cartesian coordinates $x_i$ , a distance multiplier $w_{d,i}$ , and an edge width multiplier $w_{th,i}$ . The genotype size is $4n$ , bounded in $[0,1]$ .
Graph Phenotype: An undirected spatial graph $G=(V,E)$ where edges are not explicitly encoded but inferred. An edge $(i,j)$ is considered active if a thresholded weighted distance $W_{ij} < 1$ , where $W_{ij} = (\|x_i - x_j\|/L + w_{ij})/d_{max}$ . This allows the genotype to be denser than direct adjacency encoding while spanning a sufficiently expressive subset of graph topologies.
Topology Phenotype: A planar geometric object formed by the union of active graph edges, filled with material of thickness $w_{th} \cdot th_{max}$ .

2. Multi-Objective Quality-Diversity Optimization

To navigate the vast and deceptive design space, the authors employ Pareto Dominated Novelty Search (PDNS), an extension of the Dominated Novelty Search (DNS) algorithm.

Algorithm: The method maintains a population of $N$ designs. Instead of a single fitness score, each design is evaluated on a vector of $m$ competing objectives and a behavioral descriptor $d_i \in \mathbb{R}^D$ .
Competition Fitness: A scalar competition fitness $\tilde{f}_i$ is computed based on Pareto dominance and diversity. For a solution $i$ , the set of solutions $D_i$ that Pareto-dominate it is identified. The competition fitness is the average distance in descriptor space to the $k$ -nearest dominating solutions. Solutions on the Pareto front ( $D_i = \emptyset$ ) receive infinite fitness, ensuring their preservation.
Behavioral Descriptors: A modular set of features captures geometric, topological, and parametric aspects (e.g., centroid, spread, edge density, clustering metrics, volume fraction) to guide diversity without explicit task specification.
Genetic Operators: The search uses tournament selection, node-wise two-point crossover, and a diverse set of mutations (parametric noise on position/width, and structural mutations like add/remove/split nodes).

3. Physics-Based Evaluation

Fitness evaluation utilizes high-fidelity finite element modeling (FEM) with:

Material Model: Finite strain Neo-Hookean hyperelasticity to capture non-linear material response.
Contact: The Incremental Potential Contact (IPC) method to guarantee intersection-free configurations.
Solver: Implicit Newmark temporal discretization and Newton-Raphson iteration with SparseLU factorization.
Metrics: Direct optimization of task-relevant metrics including contact forces ( $f_c$ ), contact area ( $A_c$ ), object strain energy ( $U$ ), and end-effector displacements ( $u$ ).

Experimental Setup and Results

Benchmark Validation

The graph representation was first validated on standard structural optimization benchmarks (soft gripper, inverter, and MBB beam problems). The results demonstrated that the graph-based approach successfully recovers solutions approximating known optimal topologies from traditional methods, validating its expressiveness for both compliance-based and stiffness-based problems.

Multi-Objective Soft Gripper Optimization

The framework was applied to a planar grasping task involving four distinct scenarios:

Pinch-Rigid: Small rigid object (maximize pull force).
Power-Rigid: Large rigid object (maximize contact area).
Power-Soft: Large soft object (balance strain energy and pull force).
Pinch-Soft: Small soft object (balance strain energy and pull force).

Key Findings:

Emergent Generalization: Optimization across a diverse set of grasping cases (4-objective run) produced a Pareto front with 89 non-dominated solutions. While single-objective optimizations produced "specialist" designs that dominated their specific training scenario but failed on others (overfitting), the multi-objective approach yielded designs with robust performance across all scenarios.
Morphological Strategies: Elite individuals exhibited distinct strategies (e.g., longer moment arms for large objects, hook-like structures for soft objects), yet the diverse Pareto front contained designs that could transfer effectively to unseen conditions.
Robustness to Perturbations: When tested with positional perturbations ( $\pm 10$ mm) and novel object geometries (concave column, coral-like structure), designs from the 4-objective run (and specifically the random subset of the Pareto front) maintained performance comparable to their training scenarios. In contrast, single-objective specialists showed substantial performance degradation.
Physical Validation: Limited experimental validation using 3D-printed grippers (Formlabs Elastic 50A) confirmed that the optimized designs exploit deformation and contact mechanisms predicted by simulation, with qualitative agreement in grasping trends and pull-force magnitudes.

Significance and Claims

The paper claims that diversity-driven optimization offers a principled pathway toward general-purpose soft grippers. By coupling an expressive graph-based design space with high-fidelity physics and a multi-objective quality-diversity algorithm, the authors demonstrate that:

Generalization is Emergent: Optimization over a sufficiently diverse set of grasping cases leads to designs that generalize to unseen objects and contact conditions, exhibiting improved robustness compared to task-specific solutions.
Mitigation of Overfitting: Multi-objective diversity-driven optimization mitigates the brittleness and overfitting inherent in single-objective greedy optimization.
Mechanical Intelligence: The approach leverages mechanical intelligence (compliance and adaptability) to simplify control and reduce reliance on task-specific learning, aligning with the adaptable nature of soft robotics.

The authors conclude that this framework allows robust mechanical strategies to emerge from a limited set of objectives rather than requiring scenario-specific tuning, providing a foundation for more general-purpose soft manipulation.