Skip to content
Jan Pöppel edited this page Jun 8, 2015 · 7 revisions
  • 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

Clone this wiki locally