AllegroHandDextremeADRPPO.yaml

params:
  seed: ${...seed}

  algo:
    name: a2c_continuous

  model:
    name: continuous_a2c_logstd

  network:
    name: actor_critic
    separate: False
    space:
      continuous:
        mu_activation: None
        sigma_activation: None
        mu_init:
          name: default
        sigma_init:
          name: const_initializer
          val: 0
        fixed_sigma: True

    inputs:
      dof_pos_randomized: {}
      object_pose_cam_randomized: { }
      goal_pose: { }
      goal_relative_rot_cam_randomized: { }
      last_actions: { }

    mlp:
      units: [512, 512]
      activation: elu
      d2rl: False

      initializer:
        name: default
      regularizer:
        name: None
    rnn:
        name: lstm
        units: 1024
        layers: 1
        before_mlp: True
        layer_norm: True

  load_checkpoint: ${if:${...checkpoint},True,False} # flag which sets whether to load the checkpoint
  load_path: ${...checkpoint} # path to the checkpoint to load

  config:
    name: ${resolve_default:AllegroHandADRAsymmLSTM,${....experiment}}
    full_experiment_name: ${.name}
    env_name: rlgpu
    multi_gpu: ${....multi_gpu}
    ppo: True
    mixed_precision: False
    normalize_input: True
    normalize_value: True
    value_bootstrap: False  
    num_actors: ${....task.env.numEnvs}
    reward_shaper:
      scale_value: 1.0
    normalize_advantage: True
    gamma: 0.998
    tau: 0.95
    learning_rate: 1e-4
    lr_schedule: linear #adaptive
    schedule_type: standard
    kl_threshold: 0.01
    score_to_win: 1000000
    max_epochs: ${resolve_default:1000_000,${....max_iterations}}
    save_best_after: 10000
    save_frequency: 500
    print_stats: True
    grad_norm: 1.0
    entropy_coef: 0.002
    truncate_grads: True
    e_clip: 0.2
    horizon_length: 16
    minibatch_size: 16384
    mini_epochs: 4
    critic_coef: 4
    clip_value: True
    seq_length: 16
    bound_loss_type: regularization
    bounds_loss_coef: 0.005
    zero_rnn_on_done: False

    # optimize summaries to prevent tf.event files from growing to gigabytes
    force_interval_writer: True

    central_value_config:
      minibatch_size: 16384 
      mini_epochs: 4
      learning_rate: 5e-5
      kl_threshold: 0.016
      clip_value: True
      normalize_input: True
      truncate_grads: True

      network:
        name: actor_critic
        central_value: True

        inputs:
          dof_pos: { }
          dof_vel: { }
          dof_force: { }

          object_pose: { }
          object_pose_cam_randomized: { }
          object_vels: { }

          goal_pose: { }
          goal_relative_rot: {}

          last_actions: { }

          stochastic_delay_params: { }
          affine_params: { }

          cube_random_params: {}
          hand_random_params: {}
          ft_force_torques: {}
          gravity_vec: {}
          ft_states: {}
          rot_dist: {}
          rb_forces: {}

        mlp:
          units: [1024, 512] 
          activation: elu
          d2rl: False
          initializer:
            name: default
          regularizer:
            name: None
        rnn:
          name: lstm
          units: 2048 
          layers: 1
          before_mlp: True
          layer_norm: True

    player:
      deterministic: True
      use_vecenv: True
      games_num: 1000000
      print_stats: False