dataset:
  name: ogbench/antmaze_visual
  num_steps: ${eval:'${wm.num_preds} + ${wm.history_size}'}
  frameskip: 1
  keys_to_load:
    - pixels
    - action
  keys_to_cache:
    - action