params.yaml

# Hyperparameters of PPO2
learning_params:
  cliprange: 0.2
  cliprange_vf: -1
  ent_coef: 0.003760312514286235
  gamma: 0.999
  lam: 0.8

  # Linear learning rate [LR_start, LR_end, T_start, T_end] where
  # T goes from 1 to 0 during training.
  learning_rate: [0.00025, 0.0002, 0.995, 0.2]

  n_steps: 2048
  nminibatches: 8
  noptepochs: 30

# Parameters which define the environment
environment_params:
  # Base platform parameters
  base:
    acc_mag: [0.15, 0.15, 0.3]  # Maximum accelerations [x, y, theta]
    vel_mag_lin: 0.1  # Maximum linear velocity (x^2+y^2 <= vel_mag_lin^2)
    vel_mag_ang: 0.2  # Maximum angular acceleration for theta

    # Base acceleration noise (gaussian)
    std_acc_lin: 0.001
    std_acc_ang: 0.001

  # Manipulator parameters
  joints:
    vel_mag: 0.5  # Maximum velocity for each joint
    acc_mag: 0.8  # Maximum acceleration for each joint
    base_link_name: panda_link0  # Name of the base link
    ee_link_name: panda_ee  # Link name of the end effector

    # Initial joint configuration, used for 2D lock (actuated joints will not
    # keep this value but obtain a randomly chosen angle)
    init_states: [3.141592, 1.5708, 1.5708, 0, -1.5708, 3.14159, 0.7854]

    # Names of the joints which shall be actuated
    joint_names: [panda_joint1, panda_joint4]

    # Link names and respective normalization which shall be added to the
    # observation (their 6D pose relative to base_link_name)
    link_names: []
    link_mag: []

    std_acc: 0.0001  # Acceleration noise (gaussian)

    # Position noise (uniform) for each joint during initialization. Only
    # important for fixed actuators.
    static_act_noise_mag: [0, 0, 0, 0, 0, 0, 0]

  # Reward function parameters. "Total reward" means that the reward will be
  # summed up to this value in case that the agent reaches its setpoint.
  reward:
    # Total reward for the angle between the two vectors EE and EE-to-SP
    fac_goal_dis_ang: 0

    # Reward for reducing Base-to-SP angle. Useful for differential drive.
    fac_base_sp_ang: 0

    # Harmonic potential field rewards
    rew_path_dis_p_m: 10  # Reward per meter for reducing EE-to-HPT_path dis.
    rew_path_total: 50  # Total reward for progress along HPT_path

    # Total reward for reducing Euclidean EE-to-SP distance
    fac_goal_dis_lin: 0

    # Total reward aquired for remaining in tolerance sphere "tol_lin_mag" for
    # "hold_time" seconds
    fac_sp_hold: 20

    # Additional total reward for keeping a smaller EE-to-SP distance than
    # just the required "tol_lin_mag"
    fac_sp_hold_near: 40

    # Reward for collisions
    rew_collision: -60

    # Final reward for reaching the setpoint
    rew_goal_reached: 10

    # Reward for reaching the joint limits of the manipulator
    rew_joint_limits: -20

    # Total reward for time during "timeout"
    rew_timeout: -15

    # Lidar reward parameters
    dis_lidar: 0.3  # Maximum distance at which reward will be given
    # Reward given per meter when smallest lidar value is less than "dis_lidar"
    rew_lidar_p_m: -1

  # Parameters for the different sensors of the robot
  sensors:

    # Lidar parameters
    lidar:
      ang_mag: 2.0944  # Maximum angle of the lidars (per side)

      # Link names from which scans will be calculated
      link_id1: front_laser
      link_id2: rear_laser

      n_scans: 201  # Meassurements per lidar
      range: 5  # Maximum range of the sensors

      # Noise parameters for the lidar measurements (gaussian)
      noise:
        mean: 0
        std: 0.015
        p_err: 0.025  # Probability that a scan takes maximum value ("range")

    # EE-to-SP measurement parameters
    setpoint_meas:
      # Noise for EE-to-SP measurement (gaussian)
      std_lin: 0.00025
      std_ang: 0.00025

    # Odometry parameters
    odometry:
      # Noise for odometry (gaussian)
      std_lin: 0.001
      std_ang: 0.001

  # Setpoint parameters
  setpoint:
    # Specify if reaching a SP will sometimes keep the current env and robot
    # configuration to "proceed the current mission" instead of respawning
    # the robot at a new location.
    continious_mode: true

    # Minimum time required to be in the tolerance sphere for completion
    hold_time: 1.5

    # Tolerances for the EE-to-SP vector for completion
    tol_ang_mag: 3.141592
    tol_lin_mag: 0.0425

    # Specify whether to use full3D or a 2D locked configuration. This will
    # affect if the z-coordinate is considered as well for calculating the
    # current EE-to-SP distance.
    2D_locked: true

    # Specify in which shelf layers the setpoint can spawn. The lowest layer is
    # layer 0, and the upper most layer is 3
    layers: [3]

    # Specify the maximum deviation from the center of each shelf element.
    # These are uniform distributions
    noise:
      range_x: [0.5325, 0.5325]
      range_y: [0.1, 0.19]
      range_z: [0, 0]

  # Parameters for the world
  world:
    # Uniform distributions
    corridor_length: [14, 17]  # Corridor length
    corridor_width: [2.35, 3.2]  # Corridor width
    wall_length: [0.3, 7.5]  # Length of the walls which separate doors
    wall_width: [0.05, 0.3]  # Width of all walls
    wall_height: [1.5, 1.51]  # Height of all walls
    door_length: [0.75, 2.5]  # Length of the doors
    shelf_length: [0.3, 1.75]  # Length of the spawned obstacle shelves
    shelf_height: [1.25, 1.5]  # Height of the spawned obstacle shelves
    shelf_width: [0.05, 1.5]  # # Width of the spawned obstacle shelves
    shelf_gap:  [-0.5, 5.5]  # Gap between two shelves

    # Minimum clearance in the corridor for the robot (shelf sizes will be
    # adapted to ensure this requirement)
    min_clearance: 1.25

    # Probability to spawn a completely new environment after episode end
    prob_new_env: 0.25

    # Probability to keep the robot position and environment and only spawn
    # a new setpoint
    prob_proceed: 0.25

    # Define whether the robot has an action which "breaks" the systems
    # velocities to zero or not
    use_stop_action: false

    # Harmonic Potential Field parameters
    HPF:
      # Resolution (samples per meter) specifying the discretazion of the
      # environment for creating a HPT
      res: 8

    # Maximum distance (in corridor length direction) from the setpoint
    # the robot's base platform might be spawned
    spawn_range_x: 8

    # Discretiazion resolution of the action space. When 0, the actions are
    # continous (gym.spaces.Box), when > 0 the actions are discrezited
    # (gym.spaces.MultiDiscrete)
    action_discretization: 5

    # The normalization used for the setpoint measurements.
    size: 20

    # Simulation time resolution
    tau: 0.04

    # Timeout after which the episode ends
    timeout: 120

    # Parameters for dynamic obstacles spawned (humans)
    n_humans: 2  # Number of humans which might walk around
    p_spawn_human: 0  # Probability for each human to be spawned

  # Automatic Domain Randomization parameters
  adr:
    # Thresholds for the success rate above and below which one of the
    # parameters is adapted
    success_threshold: 0.63
    fail_threshold: 0.1

    # Specifies how many episodes (per worker) are used for success rate
    # calculation. More means smoother success rate but introduces delay.
    batch_size: 5

    # Adaptions which are made to the environment in case that the success
    # rate is high enough ("success_threshold") or drops critically
    # low ("fail_threshold"). Adaptions can change environment parameters
    # over time, depending on the success rate. An example for adaptions:
    # adaptions: [
    # [
    #   {"param": "world.wall_width.1", "start": 0.01, "end": 0.3, "steps": 10},
    #   {"param": "world.p_spawn_human", "start": 0, "end": 1, "steps": 10}
    # ],
    # [
    #   {"param": "world.spawn_range_x", "start": 2.0, "end": 10.0, "steps": 10}
    # ]
    # ]
    # First, all params are initialized to their start value. Then, as soon as
    # The success rate is higher than the threshold, one of the parameters
    # p_spawn_human or wall_width[1] are adapted towards their end value in
    # a total of 10 steps. As soon as both parameters reached their end value,
    # the same procedure repeats for spawn_range_x.
    # Note that if adaptions are used, they must always be capsuled in two
    # lists, even if only a single adaption is used. The inner lists specify
    # which adaptions shall be completed before proceeding with the next
    # element in the outer list.
    adaptions: []