In-Situ Training of Anthropomorphic Robots

The control technology under development enables human operators to teach anthropomorphic robots, in the field, how to perform new complex tasks. Building upon existing inverse kinematics, rule-based control, and neural network learning technology, the training method enhances robot capabilities through operator supervision. The current innovation enables the online construction of rule-based plans of action through verbal dialogue between operator and robot, and uses verbal, visual, and manual cues such as spoken words, hand gestures, and the pushing of buttons and joysticks to teach neural networks how to improve nominal rule-based performance. Phase I results indicate that given an underlying library of intelligent behaviors, non-trivial robot task plans can be created and modified verbally by an operator in a straightforward manner. Phase II will optimize the behavior control system, training process, and operator interface. Validation will focus on two scenarios: assisting EVA astronauts with tool preparation and handling using the Johnson Space Center?s Robonaut system, and assisting a physiologically degraded astronaut in partial gravity environments. The show-and-tell approach to adaptive control is expected to give future NASA robots an unending ability to learn, and NASA astronauts the ability to customize robot behavior for both routine tasks and unexpected situations.