Problems

1. How to predict next worldstate?
   • Problem of what kind of generative model and how to incorporate the action primitives
2. How to synchronise actions and world updates?
   • World state is published with fixed frequency. How does the model know which one of all the updates is the reaction to the chosen action?
     • As long as the model finishes before the next world state is published, it is fine

Newer Problems:

• Generalization:

  • When doing pushTaskSimulation, the model only learns to start moving towards the block on a command. The actual cmdDirection is pretty much irrelevant
  • Similar in other situations. Basically the model can really only reproduce what it has seen before. Abstraction is only supported when it comes for free from the representation, i.e. global orientation and position

• Scoring problem:

  • Constants can be rather strict and do not distinguish enough yet
  • DecisionTree does not learn ideal splits (it may overfit the data as well)

• Relative coordinate system: By eliminating the orientation (block's orientation is set to 0 when going into the local coordinate system and ball orientation is fixed at 0) there is no information about the orientation left. -> Ball might drift off because the model learned that the direction changes at a certain position (correct in training when the block is turned, but wrong when the block was missed)

• Action: In the pushTask the action often only has any effect at the beginning. Even when the block turnes will the linVel be transformed so that the transformed mvDir might not even trigger any change.

• Prediction sometimes yields unreasonable configurations even with only 1 InteractionState (i.e. gripper in block) -> Figure out why/how

  • At least partially because a wrong AC is selected (i.e. only dist, dir, spos, although the block is pushed)

• Performance:

  • As the number of BaseCases increases, the performance drops, even for the AC search, (because of the current scoring method as well as the ITM-predictors)