9309189 Hodgins This is the first year of a three-year continuing award. Humans are excellent at recovering from physical failures: we can often catch ourselves when we slip on ice, trip over an electric cord, or miss a step. We can sometimes catch a dropped object and prevent it from breaking. We are not always successful in recovering from errors but present-day robots rarely even attempt such behavior. An ability to recover from errors would allow robots to operate more robustly in unpredictable and unstructured environments. This research aims to provide robots with a few simple reflexes: a tripping reflex that will allow a legged robot to recover from a stumble, a slipping reflex that will allow it to cross a slippery floor without falling, and a catching reflex that will allow it to interrupt the fall of a dropped object and prevent it from breaking. The reflexes will be designed and tested in simulation and on experimental hardware. Reflex responses are unlike most robot control algorithms because of the time constraints of a rapidly evolving dynamic system and because of the need to respond to discontinuous events. These properties make this an ideal domain for exploring questions of parameterization and choice of strategies for control. This work will help to provide a scientific foundation upon which useful transportation and manipulation systems for unstructured environments may be constructed, and perhaps to explain some behaviors observed in biological systems.
