Lalitesh K Katragadda Thesis Page

Doctoral Student, Robotics Institute
Carnegie Mellon University
Pittsburgh, PA 15213

Synergy - a language and framework for robot design

Robot design is increasingly difficult due to escalation in complexity, capability and application. A design environment can automate many design tasks and relieve the designer. Prior to robot development, designers usually compose a robot from existing or custom developed components, simulate performance, optimize configuration and parameters, and write software for the robot. Robot designers find themselves developing these facets custom to the robot using a variety of software tools from spreadsheets to C code to CAD tools. Valuable resources are expended, and very little of this expertise and development is reusable. This research considers that a language to comprehensively represent robots is lacking and the aforementioned design tasks can be automated once such a language exists. This research proposes and demonstrates the following thesis:

"A language to represent robots with a framework to generate simulation, optimize design and generate control software increases the effectiveness of design "

Synergy was prototyped and demonstrated in the context of lunar rover design, a challenging real-world problem with multiple requirements and a broad design space. Synergy was used to automatically optimize robot parameters and select parts to generate effective designs, while meeting constraints of the embedded components and sub-systems. The designs so generated are better in performance and consistency relative to designs by teams of designers using the same knowledge. Using a single representation, multiple designs are generated for four distinct lunar exploration objectives. Synergy uses the same representation to auto-generate landing simulations and simultaneously generate control software for the landing.

Synergy is composed of four software agents. A database and spreadsheet agent compiles the design and component information, generating component interconnections and ensuring consistency of types, physical units and constraints. A simulation agent generates comprehensive dynamic simulations fusing several uni-dimensional agents. An optimization agent executes rules embedded in the design, finds roots for the implicitly interconnected design and searches the parameter and component space using a genetic algorithm. A control software generator learns a generalized "neural" controller using the simulation for feedback and a genetic algorithm to guide the search.

Thesis Committee

Red Whittaker, Chair, Fredkin Professor, Director FRC, Robotics Institute, CMU

John Bares, Director National Robotics Engineering Consortium, Robotics Institute, CMU

Matt Mason , Chairman, Robotics PhD Program, Robotics Institute, CMU

Tony Stentz, Associate Director, Field Robotics Center & NREC, Robotics Institute, CMU

Robert Ambrose, Metrica Robotics, Johnson Space Center

Postscript of the submitted thesis document and presentation be accessed by anonymous ftp to cs.cmu.edu /afs/cs/user/lalit/www/thesis-web.ps. See the original proposal for reference and evolution of the idea.