Packbot Rollover Prevention
We applied Dynamic Threshold Learning to learn how to prevent the PackBot from rolling over when climbing obstacles. In these experiments, DTL learns the combination of pitch angle (state) and velocity (action) that causes the robot to roll over when climbing over a tall obstacle. After learning this state-action map, DTL is able to serve as a guarded teleoperation behavior, preventing the robot from rolling over under operator control. This video shows DTL learning to prevent PackBot rollover:
After a few learning trials, DTL is able to prevent rollover for a wide range of speeds. Additional details are available in our paper submitted to ICRA 2012.
