push-DQN_config.yaml

model:
  path: #/home/baris/Workspace/push-to-see/saved_models/push-DQN/
  file: new #snapshot-41k_CoRL.pth # set 'new' for not loading a snapshot of a pretrained model

detection_thresholds:
  confidence_threshold: 0.85 # to define whether a prediction from mask_rg will be considered as a possible object/a detection
  mask_threshold: 0.85 # to define whether a detection will be considered correct based on its quality (i.e. how well its segmentation mask matches the ground truth)
  session_success_threshold: 0.90 #

logging:
  path: Default # Set 'Default' for ./log
  continue_logging: False # to continue logging from previous session #TODO
  save_visualizations: True # to save visualisations of FCN predictions
  snapshot_saving_rate: 100 # to define how often the model weights will be saved

environment:
  workspace_limits: [[-0.724, -0.276], [-0.224, 0.224], [-0.0001, 0.4]] # to define workspace limits in robot coordinates (Cols: min max, Rows: x y z)
                                                                        # (default val [[-0.724, -0.276], [-0.224, 0.224], [-0.0001, 0.4]])
  heightmap_resolution: 0.002 # meters per pixel of heightmap (default val 0.002)

  min_num_objects: 26 # Minimum number of objects that will be dropped in the simulation
  max_num_objects: 32 # Maximum number of objects that will be dropped in the simulation (LIMIT 72!) #TODO current limit is less than 72 --> change the colour space of printed mask diff
                      # The total number of objects per scene will be randomised between min and max!
  session_limit: 30 # The number of iterations in each session. If a task cannot be solved within this limit, it's considered as a fail, otherwise success

setup:
  is_testing: False # If True --> Exploration Probability will be 0.0 !!!
  exploration_probability: 0.5 # Initial exploration probability
  epsilon_greedy_policy:
    exploration_rate_decay: True # to set epsilon greedy exploration (from initial probability to min_exp_rate_decay)
    min_exp_rate_decay: 0.1 # to define minimum value to which the exploration probability can reach
  is_random_model: False # If true, exploration probability is set to 1.0 with no rate decay (CAUTION! if true, overrides above values and generates a total random model!)

training:
  future_reward_discount: 0.5