Recent developments in the separate domains of robot force control and neurocontrol have advanced to the point where integration of these technologies may yield a dramatic leap forward in robot competence. At present, robot contact operations are clumsy and ineffective, except in instances where the environment is highly predictable and the tasks can be accomplished using position control alone. However, for futuristic applications of robots, such as autonomous assembly of a space-based power system, robots for defense applications, or a potential industry in domestic robotics, competence in contact operations in unstructured environments is essential for success.
The exploratory project will evaluate the use of intelligent control algorithms with appropriate set of assumptions and representations of robot/environment contact dynamics, using "Natural Admittance Control" and "virtual dynamics" as a means to integrate high and low-level controls.
It will consider four tasks in intelligent control research: (1) constructing a reflexive damage-avoidance layer to assure the safety of a robot and its environment and enable the capability for safe, autonomous learning; (2) human strategies for contact operations will be analyzed and "cloned" for robot execution, using identification of atomic behaviors and recognition of events that trigger sequential behaviors; (3) on-line identification of environment dynamics; (4) neural-net based optimizing control algorithms employing "adaptive critics" will be constructed in the context of NAC/virtual dynamics.
