Abstract
In the robotics community, multiple robots can be connected to form a serially connected robot, enabling adaptation to diverse environments and task requirements. However, the significantly large design space, high-dimensional observations, and complex actions pose significant challenges in designing serially connected robots with well-adapted morphologies. To address these problems, we propose a multi-objective optimization framework that first optimizes a single robot for easy tasks, then replicates its structure to construct a serially connected robot for hard tasks. For generalizing across scale variations of a single morphology, we propose a novel control method based on information sharing to learn a two-robot controller with combinatorial generalization capabilities. The two-robot controller can zero-shot generalize various sizes of serially connected robots, ensuring good motion coordination while significantly reducing the cost of control optimization. Experimental results show that the proposed algorithm outperforms the other design algorithm in most tasks. The serially connected robots can achieve reliable movement. By combining information sharing with the two-robot controller, agile motion can be achieved under complex terrain conditions. Compared to baseline algorithms, the proposed algorithm achieves superior performance on all hard tasks, especially with large-scale serially connected robots.
Introduction
Designing and controlling chains of modular soft robots is notoriously difficult due to the "curse of dimensionality" and complex physical interference between units. While connecting multiple simple robots into a "serially connected" chain—akin to a robotic centipede—allows systems to tackle challenges a single unit cannot, such as crossing wide gaps or climbing steep slopes, two major hurdles exist. First, the search space for the optimal arrangement of soft and rigid materials (voxels) is astronomical. Second, as more robots are added to a chain, the number of sensors and actuators grows, making traditional Reinforcement Learning (RL) struggle to learn effective, synchronized policies. This paper introduces the Observation Sharing Multi-Objective Evolutionary Algorithm (OS-MOEA), a framework that co-optimizes robot morphology and control to enable scalable, coordinated motion in large-scale chains.
Methodology
Voxel-Based Design and Representation
The proposed framework utilizes a voxel-based representation where robots are modeled as a grid of cells. Each voxel can be assigned one of five states: empty, rigid, soft, or an actuator (horizontal or vertical). This discrete representation allows for a direct mapping between a numerical "material matrix" and the physical structure of the robot, facilitating both single-unit design and the assembly of serially connected chains.
OS-MOEA Framework
The OS-MOEA framework addresses design complexity through a two-level multi-objective optimization process using NSGA-II. It treats the performance of a single robot on an "easy" task (e.g., flat-ground locomotion) and the performance of a multi-robot chain on a "hard" task (e.g., gap jumping) as separate objectives. This ensures that the basic building block is functional before it is scaled into a larger system.
Information Sharing Control
To overcome the control bottleneck, the authors propose a "two-robot" training strategy. Instead of training a controller for a full 10-robot chain, a controller is trained for a simple two-unit system.
Crucially, these units share velocity observations during training, creating a "global" awareness that facilitates synchronization. This learned controller can then be applied pairwise across a chain of any length, enabling zero-shot generalization to large-scale robots without additional training. This "Power of Two" approach bypasses the dimensionality disaster of trying to train a single neural network to control a massive chain.
Experimental Results
The framework was evaluated using the EvoGym benchmark across several tasks, including object manipulation (Carrier-v0) and complex terrain navigation (BridgeWalker-v0, GapJumper-v0).
Scalability and Performance
OS-MOEA demonstrated superior scalability compared to baseline algorithms. In the "BridgeWalker" task, the framework maintained high performance even as the number of connected robots ($N_c$) increased from 2 to 10. While baseline performance plummeted as complexity grew (dropping from 4.86 to 1.87), OS-MOEA's rewards remained stable, dropping only slightly from 5.23 to 4.65.
The Role of Information Sharing
Ablation studies confirmed that information sharing is critical for coordination. Without it, robots in a chain act asynchronously, leading to structural instability and poor performance. By sharing average velocity between adjacent units, the robots naturally align their motion states, achieving agile movement even under complex terrain conditions.
Conclusion
The OS-MOEA framework provides a robust solution for the design and control of serially connected soft robots. By combining multi-objective evolutionary optimization with a scalable information-sharing control policy, the system bypasses the dimensionality disaster typically associated with large-scale modular robotics. Future work will focus on bridging the sim-to-real gap and developing universal controllers capable of generalizing across diverse robot morphologies.
Summary updated