config.yaml

resume: false
device: cuda
use_amp: true
seed: 1
dataset_repo_id: lerobot/pusht_keypoints
video_backend: pyav
training:
  offline_steps: 0
  num_workers: 12
  batch_size: 256
  eval_freq: 10000
  log_freq: 250
  save_checkpoint: true
  save_freq: 10000
  online_steps: 1000000
  online_steps_between_rollouts: 1000
  online_rollout_n_episodes: 10
  online_rollout_batch_size: 10
  online_sampling_ratio: 1.0
  online_env_seed: 10000
  online_buffer_capacity: 40000
  online_buffer_seed_size: 0
  do_online_rollout_async: false
  image_transforms:
    enable: false
    max_num_transforms: 3
    random_order: false
    brightness:
      weight: 1
      min_max:
      - 0.8
      - 1.2
    contrast:
      weight: 1
      min_max:
      - 0.8
      - 1.2
    saturation:
      weight: 1
      min_max:
      - 0.5
      - 1.5
    hue:
      weight: 1
      min_max:
      - -0.05
      - 0.05
    sharpness:
      weight: 1
      min_max:
      - 0.8
      - 1.2
  grad_clip_norm: 10.0
  lr: 0.0003
  delta_timestamps:
    observation.environment_state:
    - 0.0
    - 0.1
    - 0.2
    - 0.3
    - 0.4
    - 0.5
    observation.state:
    - 0.0
    - 0.1
    - 0.2
    - 0.3
    - 0.4
    - 0.5
    action:
    - 0.0
    - 0.1
    - 0.2
    - 0.3
    - 0.4
    next.reward:
    - 0.0
    - 0.1
    - 0.2
    - 0.3
    - 0.4
eval:
  n_episodes: 50
  batch_size: 50
  use_async_envs: true
wandb:
  enable: true
  disable_artifact: true
  project: lerobot
  notes: ''
fps: 10
env:
  name: pusht
  task: PushT-v0
  image_size: 96
  state_dim: 2
  action_dim: 2
  fps: ${fps}
  episode_length: 300
  gym:
    obs_type: environment_state_agent_pos
    render_mode: rgb_array
    visualization_width: 384
    visualization_height: 384
policy:
  name: tdmpc
  pretrained_model_path: null
  n_action_repeats: 1
  horizon: 5
  n_action_steps: 5
  input_shapes:
    observation.environment_state:
    - 16
    observation.state:
    - ${env.state_dim}
  output_shapes:
    action:
    - ${env.action_dim}
  input_normalization_modes:
    observation.environment_state: min_max
    observation.state: min_max
  output_normalization_modes:
    action: min_max
  image_encoder_hidden_dim: 32
  state_encoder_hidden_dim: 256
  latent_dim: 50
  q_ensemble_size: 5
  mlp_dim: 512
  discount: 0.98
  use_mpc: true
  cem_iterations: 6
  max_std: 2.0
  min_std: 0.05
  n_gaussian_samples: 512
  n_pi_samples: 51
  uncertainty_regularizer_coeff: 1.0
  n_elites: 50
  elite_weighting_temperature: 0.5
  gaussian_mean_momentum: 0.1
  max_random_shift_ratio: 0.0476
  reward_coeff: 0.5
  expectile_weight: 0.9
  value_coeff: 0.1
  consistency_coeff: 20.0
  advantage_scaling: 3.0
  pi_coeff: 0.5
  temporal_decay_coeff: 0.5
  target_model_momentum: 0.995