Olac Fuentes and Randal Nelson 5. Conclusions and Future Work 4. Related Work

interests are in the elds of robotics and machine learning, including dextrous manipulation, teleoperation, autonomous mobile robots and adaptive robot behavior. LEARNING DEXTROUS MANIPULATION SKILLS 15 Heuristics derived form observations made on human hands were used to reduce the degrees of freedom of dextrous manipulation with robotic hands. This signiicantly simpliied the task and made autonomous learning feasible. Our system does not rely on simulation. Instead, all the experimentation is done by a physical robot. This is valuable in situations such as dextrous manipulation , where building a realistic and accurate simulator is extremely diicult. We used a modiied version of the evolution strategy to learn manipulation primitives. This learning algorithm successfully dealt with the noise in sensors and eeectors and allowed the primitives to be learned in a period of a few minutes. We showed that the learned primitives can be combined to form general manipulations and perform more complex tasks. Future work includes a quantitative comparison between our approach and more traditional non-learning approaches to dextrous manipulation. It also includes learning primitives that require repositioning the ngers on the surface of the object , and using a more sophisticated version of the evolution strategy to learn the primitive skills in even shorter periods of time. 6. Acknowledgements We would like to thank the anonymous reviewers for their helpful comments and suggestions. We would also like to thank Chris Brown, Dana Ballard, Roger Gans and Martin JJ agersand for several interesting discussions. Notes 1. Some modiications to the original binary-valued representation have been proposed and used somewhat successfully 10] 2. In the implementation we use a xed length window of past results to estimate this probability 3. This observation has been used in other systems (e. g. 19]) for computing the inverse kine-matics of the Utah/MIT hand. References 1. Michael Arbib, Thea Iberall, and Damian Lyons. Coordinated control programs for movements of the hand. Dunn and Segen 6] presented a robotic system that learns how to grasp objects. In their system, when an object is presented for the rst time the robot experiments with it, seeking a way to grasp it by trial and error using visual information and input from the robot gripper. A discovered grasp is saved along with the object's shape. The system generalizes to diierent positions and orientations but not to sizes. Kamon et al. 14] presented a robotic system that learned to grasp objects …