Encoding safe task achievement in humanoid robots and autonomous agents
Intelligent Robotics Lab
University of Texas at Austin
How can a humanoid robot safely walk on uneven ground, without having prior models of its environment? How can an autonomous trading agent operate safely in an unknown random environment? I will show that it is possible to design such strategies using two components: (a) task encoding in terms of qualitative models and (b) techniques for learning, search and optimization to enable adaptation to the environment. Qualitative models ( i.e., simple abstractions whose phase space geometry may be used to specify spatial, temporal and dynamical constraints) enable principled analysis of safety and structural stability. Machine learning allows the agent to bind these abstractions to the more complex environment the agent actually interacts with. This allows the agent to deal effectively with imprecision and environmental variation. I will conclude with a brief discussion on how this methodology and representation enables new algorithms for behavior discovery and verification.