Our question
#2
by bay-llm - opened
The paper states that state-action sequences were recorded at 20 Hz and multi-view images were rendered by rolling out the recorded actions, while the released LeRobot dataset has fps=10.
Could you clarify:
- how the 20 Hz state-action sequence was converted to the released 10 Hz dataset;
- whether each released row is (observation_t, action_t) before applying action_t, or (observation_{t+1}, action_t) after applying it;
- whether actions were subsampled, accumulated, averaged, or otherwise transformed;
- at which simulator step the rendered images/state are captured relative to the stored action?