This project introduces a new concept of Intelligent Selective Passivity (ISP). Whereby, a robot acts passively when it fails to perform well in a task and non-passive otherwise. Passivity is achieved by embedding a virtual energy tank inside a reinforcement learning loop and ISP is achieved by refilling the energy tank in proportion to a Potential-based reward function defined for the task.
By using this approach firstly, the energy in the tank can be limited within a safety limit always i.e. in learning mode for a task as well in an inference mode when the controller tuned using reinforcement learning acts as a state feedback controller. Secondly, the robot can learn to avoid the energy tank singularity by acting energy efficiently and showing high task proficiency. Finally, no simulation is needed to initialize the energy tank for a complex task.
A proof of concept for this approach is presented. 2 experiments showing that a dynamic energy tank can make the robot truly energy aware and aid the learning process is also presented.
To join the presentation via Microsoft Teams click here