|
Abstract
|
Reinforcement learning "in-situ" avoids the so-called reality gap. In the meanwhile, the trained agent incurs hazards (mechanical fatigue or accidents).
The proposed developmental RL approach is remotely inspired by the learning of motion primitives by infants. Natural limitations prevent infants from endangering their body when exploring their sensori-motor abilities. As children grow, their musculoskeletal system evolves and becomes stronger. Adults, who can operate with full strength, are able to seriously injure themselves; but they are assumedly skilled enough to avoid injuries with high probability.
The PhD defines a robotic developmental process by gradually extending the actuator ranges. External rewards are used to determine when and how the robot actuator range can be extended. Then, the intensity of the robot actions will increase as its skills improve.
The process will be iterated until the robot is able to deploy the desired skills with full strength.
|