seed: 1
record_video: True    # Is considered False if <environment.render> = False below

environment:
  render: True
  num_envs: 1200
  eval_every_n: 1000
  save_every_n: 50
  num_threads: 'auto'
  simulation_dt: 0.0025
  control_dt: 0.1
  max_time: 24.0
  log_interval: 8
  enable_dynamics_randomization: False
  episode_length_for_dynamics_learning: 2.0

  action_scaling: 1.0

  velocity_command:
    limit_heading_velocity: 1.5
    limit_lateral_velocity: 1.0
    limit_yaw_rate: 1.5

    # Less than this is considered to be a zero velocity command
    limit_velocity_magnitude: 0.35

    # These values are rescaled if their sum is not equal to 1
    probability_zero_command: 0.2
    probability_heading_command: 0.0
    probability_lateral_command: 0.0
    probability_yaw_command: 0.0
    probability_direction_command: 0.0
    probability_constant_command: 0.8   # includes all three velocities

    # Time range in seconds when the next command is sampled
    command_sampling_time_min: 3.0
    command_sampling_time_max: 4.0

  reward:
    base_orientation:
      coeff: -5.0
    base_linear_velocity_tracking:
      coeff: 2.5
    base_angular_velocity_tracking:
      coeff: 1.25
    joint_torque:
      coeff: -0.025
    joint_velocity:
      coeff: -0.0
    joint_position:
      coeff: -0.25
    action_smoothness:
      coeff: -0.1
    feet_clearance:
      coeff: 2.25
    feet_slip:
      coeff: -0.25
    pronking:
      coeff: -1.0
    base_height:
      coeff: -0.
    symmetry_z:
      coeff: -0.0
    feet_deviation:
      coeff: -0.0
    trotting:
      coeff: 0.0
    joint_jerk:
      coeff: 0.0
    vertical_linear_velocity:
      coeff: -2.0
    horizontal_angular_velocity:
      coeff: -0.05

  observation_indices:
    rotation: [ 0, 3 ]
    joint_position: [ 3, 15 ]
    angular_velocity: [ 15, 18 ]
    joint_velocity: [ 18, 30 ]
    linear_velocity: [ 30, 33 ]
    command: [33, 36]
    joint_position_error: [36, 48]

    positions: [ 0, 15 ]
    velocities: [ 15, 33 ]

    actor_input: [ 0, 48 ]
    critic_input: [ 0, 48 ]

    recurrent_input: [0, 48]

    dynamics_encoding_input: [ 0, 33 ]
    dynamics_inference_output: [ 0, 33 ]

    synchronous_estimation: None

  curriculum:
    reward_factor: 1.0
    reward_advance_rate: 1.0

module:
  # Choose from ['dense', 'recurrent', 'dynamics_encoding']
  type: 'dense'

  actor:
    # Choose from ['tanh', 'leaky_relu', 'softsign']
    activation: 'tanh'
    hidden: [256, 256]

  critic:
    # Choose from ['tanh', 'leaky_relu', 'softsign']
    activation: 'tanh'
    hidden: [256, 256]

  properties:
    dense:
      shuffle_batch: True
      predict_values_during_act: False

      initial_action_std: 1.0
      compute_jacobian: True

      network_weights_gain: 1.0

    recurrent:
      shuffle_batch: False
      predict_values_during_act: True

      hidden_state_includes_state_embedding: False

      initial_action_std: 1.0

      hidden_state_dim: 64

      fast_evaluation: False
      fast_evaluation_over_full_batch: True

      switch_to_slow_evaluation: False
      switch_to_slow_evaluation_steps: 0

      critic_grad_flow: True
      compute_jacobian: True

      network_weights_gain: 1.0

    dynamics_encoding:
      latent_dynamics_includes_state_embedding: False

      infer_accelerations_only: True

      initial_action_std: 1.0

      shuffle_batch: False
      predict_values_during_act: True

      hidden_state_dim: 64
      latent_dynamics_dim: 64

      fast_evaluation: True
      fast_evaluation_over_full_batch: True

      synchronous_update: True

      critic_grad_flow: True
      dynamics_inference_delay: 4
      compute_jacobian: True

      enable_early_termination_for_dynamics_learning: False

      network_weights_gain: 1.0

      # If greater than 1, the encoding and decoding blocks update <rl_update_decimation> - 1 times before RL update
      rl_update_decimation: 1
      rl_update_decimation_decay_rate: 1
      rl_update_decimation_min: 1

algorithm:
  update_steps: 8
  gamma_half_life_duration: 3.0 # seconds

  learning_rate:
    initial: 5e-4
    final: 5e-4

    min: 5e-4
    decay_steps: 1

    mode: 'constant'