# @package _global_

# Use `act_real_no_state.yaml` to train on real-world Aloha/Aloha2 datasets when cameras are moving (e.g. wrist cameras)
# Compared to `act_real.yaml`, it is camera only and does not use the state as input which is vector of robot joint positions.
# We validated experimentaly that not using state reaches better success rate. Our hypothesis is that `act_real.yaml` might
# overfits to the state, because the images are more complex to learn from since they are moving.
#
# Example of usage for training:
# ```bash
# python lerobot/scripts/train.py \
#   policy=act_real_no_state \
#   env=dora_aloha_real
# ```

seed: 1000
dataset_repo_id: lerobot/aloha_static_vinh_cup

override_dataset_stats:
  observation.images.cam_right_wrist:
    # stats from imagenet, since we use a pretrained vision model
    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
  observation.images.cam_left_wrist:
    # stats from imagenet, since we use a pretrained vision model
    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
  observation.images.cam_high:
    # stats from imagenet, since we use a pretrained vision model
    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
  observation.images.cam_low:
    # stats from imagenet, since we use a pretrained vision model
    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)

training:
  offline_steps: 100000
  online_steps: 0
  eval_freq: -1
  save_freq: 20000
  save_checkpoint: true

  batch_size: 8
  lr: 1e-5
  lr_backbone: 1e-5
  weight_decay: 1e-4
  grad_clip_norm: 10
  online_steps_between_rollouts: 1

  delta_timestamps:
    action: "[i / ${fps} for i in range(${policy.chunk_size})]"

eval:
  n_episodes: 50
  batch_size: 50

# See `configuration_act.py` for more details.
policy:
  name: act

  # Input / output structure.
  n_obs_steps: 1
  chunk_size: 100 # chunk_size
  n_action_steps: 100

  input_shapes:
    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
    observation.images.cam_right_wrist: [3, 480, 640]
    observation.images.cam_left_wrist: [3, 480, 640]
    observation.images.cam_high: [3, 480, 640]
    observation.images.cam_low: [3, 480, 640]
  output_shapes:
    action: ["${env.action_dim}"]

  # Normalization / Unnormalization
  input_normalization_modes:
    observation.images.cam_right_wrist: mean_std
    observation.images.cam_left_wrist: mean_std
    observation.images.cam_high: mean_std
    observation.images.cam_low: mean_std
  output_normalization_modes:
    action: mean_std

  # Architecture.
  # Vision backbone.
  vision_backbone: resnet18
  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
  replace_final_stride_with_dilation: false
  # Transformer layers.
  pre_norm: false
  dim_model: 512
  n_heads: 8
  dim_feedforward: 3200
  feedforward_activation: relu
  n_encoder_layers: 4
  # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
  n_decoder_layers: 1
  # VAE.
  use_vae: true
  latent_dim: 32
  n_vae_encoder_layers: 4

  # Inference.
  temporal_ensemble_coeff: null

  # Training and loss computation.
  dropout: 0.1
  kl_weight: 10.0