realrobot_image.yaml

# @package _global_
cross_validate: False
save_intermedia_models: True

dataset_generator_func:
  _target_: env.realworld.rollout_runner.convert_dataset_image
  episode_num_pertask: 10000 
  env_names: ["RealRobotDrakeHammerforhammerEnv-Tool"]

# test
rollout_runner:
  _target_: env.realworld.rollout_runner.RolloutRunner
  episode_num: 50
  env_names: ["DrakeHammerEnv"]
  save_video: True

dataloader:
  batch_size: 256 
  num_workers: 4
  pin_memory: True
  persistent_workers: False

val_dataloader:
  batch_size: 256

dataset:
  _target_: hpt.dataset.local_traj_dataset.LocalTrajDataset
  use_multiview: True
  observation_horizon: 2
  use_disk: True

head:
  _target_: hpt.models.policy_head.MLP
  input_dim: ${network.embed_dim}
  tanh_end: True # normalized action output
  output_dim: -1 # overwrite based on dataset
  widths: [256, 128]
  normalize_action: ${head.tanh_end}
  dropout: True

train:
  total_epochs: 3000  # maximum training epochs before termination
  total_iters: 20000 # maximum training steps before termination
  epoch_iters: 1000  # training steps in each epoch
  validation_iters: 100 # maximum iterations for validation


domains: real_robot_image