Optimal Computational Design Framework for Modular Manipulators

Modular manipulators, composed of multiple independent modules, offer a versatile solution to meet the growing need for mass-customized manipulation tasks. These robots can be swiftly reconfigured into various morphologies by altering the interconnections between the modules. This flexibility allows them to adapt seamlessly to a wide array of tasks and environments. The capability to rearrange these modules into diverse configurations enhances the robot’s adaptability, making it highly effective across different manipulation scenarios. To effectively deploy the modular manipulator in various environments for specific tasks, it is essential to develop a computational framework that optimizes both the manipulator's mounted pose and its morphology design. The performance of the modular manipulator in different settings is critically dependent on these factors. By tailoring the mounted pose and morphology to the task and environment, the manipulator can achieve optimal functionality and adaptability. In this work, we introduce a comprehensive computational framework designed to optimize both the physical morphology and mounted pose of modular manipulators, tailored for specific tasks and execution environments. This framework is structured around three principal components: task-trajectory generation, computational design, and control of the modular manipulator. (i) Task-Trajectory Generation: This component defines manipulation tasks as trajectories within Cartesian space, characterized by sequences of desired end-effector poses. The underlying assumption is that the successful execution of these task-specific trajectories signifies the task's achievability. (ii) Control for modular manipulator: Hierarchical Model Predictive Control (H-MPC) is employed as the high-level controller to manage the execution of these trajectories, dynamically adjusting the manipulator's morphology to navigate the specified paths effectively. The performance of these executions, governed by the H-MPC, is continuously assessed using specific evaluation metrics that influence the ongoing optimization process. These metrics assess various morphologies and mounted poses, refining the control strategies based on real-time data and performance evaluations. The control component incorporates environmental information to enhance adaptability and employs the MPC framework to compute kinematic solutions that address both kinematic and dynamic constraints. (iii) Computational Design: This component optimizes the robot's physical configuration. Unlike traditional methods, our strategy introduces a novel mapping function that \textit{implicitly} represents the manipulator's morphology, transforming its discontinuous morphological states into a continuous optimization space. This allows for a seamless integration of the design and mounted pose into the continuous domain, facilitating a more robust search for optimal configurations. We utilize a sample-based optimization method, particularly effective in handling the complexities of mixed-integer optimization problems, to find high-performance solutions efficiently. Following the proposed computational design framework for optimizing the modular manipulator in the single-branch manipulator, when a task requires a larger workspace, extending the manipulator’s length becomes necessary. This extension increases the torque demands on the joints near the base, posing a challenge in managing these forces without exceeding torque limits. To address this, we propose a bi-branch design that adds an additional branch to the traditional single-chain manipulator. This approach eliminates the need to redesign the robot or upscale the actuators used in the modules. This design reduces the torque on the base joints by allowing the main branch to focus on task execution, while the assist branch counteracts external forces at the shared link. Although the bi-branch design has the potential to reduce joint loads, its effectiveness depends on the specific task. To optimize the design of modular manipulators in both single-branch and bi-branch configurations, we introduce a computational framework that derives the optimal design from motion planning results. We validated this framework by optimizing the manipulator’s morphology with various initial guesses and demonstrated its practical effectiveness through experiments with the bi-branch design, performing drilling tasks on the CONCERT modular robotic platform, which outperformed prior designs.