Notes

• Machine is called "ravenwood"

• Prediction seems to generalize a lot better when predicting differences instead of target values!!

• "Ignore" the complete worldstate. Store cases only for object states as well as interaction states.

• When predicting a new world state:

  • Search for best matching OS/IS and predict those for each OS/IS in the current world state.
  • Prediction score will be evaluated on each element seperately as well.

• Using constants to reduce ITM feature vector improves pushTaskSimulation performance in the sense that predicting the gripper position and the block are more aligned -> less often will the gripper be predicted to be INSIDE the block

---

• Transition probabilities for the ACs does not improve selection!
  • Try taking cmd into account