Automated Task-Based Synthesis and Optimization of Field Robots

We present Darwin2K, a widely-applicable, extensible software tool for synthesizing and optimizing robot configurations. The system uses an evolutionary optimization algorithm that is independent of task, metrics, and type of robot, enabling the system to address a wide range of synthesis problems. Darwin2K can synthesize fixed-base and mobile robots (including free-flying robots, mobile manipulators, modular robots, and multiple or bifurcated manipulators), and includes a toolkit of simulation and analysis algorithms which are useful for many synthesis tasks. Some of these capabilities, such as dynamic simulation, are novel in automated synthesis of robots. An extensible software architecture enables new synthesis tasks to be addressed while maximizing use of existing system capabilities; this extensibility is a key contribution of the system. A key challenge is effectively optimizing multiple performance metrics; we present a method called Requirement Prioritization that guides the evolutionary algorithm through the design space. We apply Darwin2K to a robot synthesis tasks that includes synthesis of robot kinematics, dynamics, structural geometry, and actuator selection to meet and optimize multiple performance requirements. Introduction: Why Automated Synthesis? Robot configuration design is characterized by generating an artifact which is capable of motion in an area or volume. Typically, the configuration process generates the overall form of the robot, including kinematics and other geometry at the bare minimum but often including approximate descriptions of inertial properties, actuator and material selection, and structural geometry. This design process is often performed in an ad hoc manner. It can be difficult to translate the requirements of a task into a robot configuration, and in contrast to some other engineering disciplines there are few design rules that can simplify this process. Human designers rely on intuition and experience with related design problems, and on engineering rules based on the experience of other expert designers. When designing a robot for a task with many new characteristics, relevant experience may be quite limited and may unnecessarily restrict the range of designs that are explored. Frequently, a human investigates a small number of concepts on paper and selects a few that look promising. More detailed studies may then be performed on these, culminating in the simulation of one or more designs. One of the candidates is selected for detailed design, with tools such as finite element analysis used to evaluate parts of the design. Once the robot is built, changes may be required due to unforeseen problems or interactions s that were not modeled in simulation. Significant design iterations are often not practical, since much of a project’s schedule and resources may be devoted to creating a single robot; building a second or third robot to remedy design flaws is out of the question for many robot design projects. Thus, it is crucial to perform as much analysis and simulation as possible before the robot is built, and it is highly desirable to “get it right” the first time -since the first time may be the only time. Because of these factors, automated synthesis tools are especially attractive for robot design. Synthesis tools can address design problems for which there are little or no relevant human experience, can explore much larger numbers and ranges of designs than a human, can quickly perform design iterations in simulation, and can produce a well-optimized solution with high confidence of performance, all of which contribute to the likelihood of success of the first physical implementation. Darwin2K is a software toolkit for automated synthesis of robot configurations. It includes capabilities for quickly describing and modifying robot configurations, simulating configurations as they perform tasks, and automatically synthesizing configurations to meet task-specific requirements and to optimize performance. Darwin2K is very useful in the early stages of the configuration process, as it can automatically explore tens of thousands of designs and can allow a human designer to rapidly perform design iterations and evaluate potential robots in simulation.