Automation Tool for Rapid Design of Space Systems

Designing spacecraft and planetary exploration devices involves a lengthy and painstaking process of trying and comparing very large numbers of possible combinations of designs and components. We engaged in this study to explore the possibility that this process could be automated to a significant extent, offering mission architects and engineers an interactive tool that would enable consideration of many more designs than could reasonably be addressed manually, and saving large amounts of time and money. In particular, we sought to demonstrate a systematic methodology for an analytical framework that would provide design teams with a better basis upon which to make design decisions in both the early conceptual and the later developmental phases of a space mission.

As a case study for our approach, we investigated how to automate the design of the MER manipulator. The handcrafted design which is flying on the MER rovers is shown here.

We took as our problem the robotic manipulator arm, known as the Instrument Deployment Device (IDD), for the MER rovers that are scheduled to explore Mars in 2004. Our goal was to develop an automated system that could produce a set of designs, using commercially available parts, that would approximate the functionality and constraints of the design that was actually handcrafted for the MERs, but using far fewer workhours.

Process

We began with science and engineering requirements derived from the MER mission requirements. The rover arm needed to be capable of carrying and manipulating a turret holding four instruments: a rock abrasion tool, a micro-imager, an alpha particle x-ray spectrometer (APXS), and a M�ssbauer spectrometer. The arm had to be able to position each of the tools against any surface within a defined workspace, with a precision and contact pressure specific to each tool.

We gave the algorithm a "toolbox" of off-the-shelf components, and a set of requirements derived from mission requirements and desired science return. The primary design drivers were: (1) at least 5 degrees of freedom (representing the arm's ability to move its instruments through different planes), (2) minimizing mass, (3) maximizing the accessible dexterous workspace (the volume of space in which the arm could properly position each of the four instruments in its turret), and (4) stowage volume (ability to be stowed in a compartment of a certain size).

We constrained the enormous number of possible combinations (the trade space or design space) by telling the algorithm that it could ignore certain potential design elements that we knew from experience would not be suitable (no prismatic joints, for example), but still left quite a formidable trade space of some 8x1016 unique configurations for the algorithm to explore. Taking one design element as an example -- "right angle joint 2" -- there were 135,168 possible combinations of motors, gearboxes, materials, diameters, wall thicknesses, and overall lengths to consider.

Our algorithm was built upon Darwin2K, a design system developed at Carnegie Mellon University by Chris Leger. It uses a genetic algorithm approach to searching the trade space. In an analogy to genetic mutation, the algorithm periodically shifts its search methods (except when the current method proves increasingly productive) to enable it to cover a wider variety of trade space regions than time would permit if it exhausted one search method before beginning the next. We ran the algorithm on a network of five Unix stations.

We developed a "seed design" for the algorithm to use as a starting point, which conveyed in broad strokes the general form we expected the arm to take.

As the algorithm ran, it went through a continuous process of assembling components from its "toolbox" database, comparing them to our constraints and requirements, and refining the design. It evaluated each candidate design according to the four metrics considered the most important: dexterous workspace, position accuracy, link deflection (how much the link bends under the load of the instruments at the tip of the arm), and collision prevention (with itself, the rest of the rover, and the environment). Designs that failed to meet all of those requirements were rejected. When designs passed that hurdle, the algorithm attempted to improve them with regard to their mass, power consumption, and time requirements.

Four of the resulting designs. The "mass metric" design (large image) was considered one of the best overall.

After the algorithm evaluated 60,000 configurations (a run time of 40 hours), it submitted the best design in each of the above seven metrics.

A human engineer carefully evaluated each of the designs, and selected one to use as the seed design for the next run. He had the opportunity to make manual modifications at this point, and in fact did so after the fourth run. This step also offered the possibility of modifying constraints and requirements for the next run. We performed a total of 10 such runs, each time using the best result of the previous run as the starting point for the next one.

The results of four selected runs are shown in the following graph:

In this star plot, the center represents the most desirable results. The red line represents the handcrafted MER IDD, and is used for comparison and validation of the approach.

This graph illustrates the fact that improvements in one metric often come at the expense of others. However, design samples 3 and 4, which were among the last of the 10 actually produced, show the best overall results in the most important metrics.

Results

While not fully equaling the performance of the handcrafted arm at this point, ten iterations of the design process produced results that were not far off, and would provide an excellent head start for a human designer. We estimate that this process used approximately one-fifth as many workhours as a fully handcrafted design would have taken to reach a comparable result. At the same time, it enabled exploration of more design architectures than might otherwise be considered during the conceptual design phase.

This analytical framework provides a highly comprehensive exploration of the system tradeoff space, and enables designers to focus their efforts in the regions most likely to yield the best, most cost-effective design solutions. We believe that it demonstrates the advantages of combining the number-crunching power of a good algorithm with the experience, intuition, and judgment of a good human engineer. It holds the promise of making the design process much more efficient at a great savings in cost and time.