Parameters

All parameters are defined in args.py. This page provides an overview of the most important parameters and their applicability to each workflow step.

Parameter Overview

The following table lists the key parameters and indicates which workflow steps they apply to:

Parameter	Description	Data Gen	Train	Open Loop	Closed Loop
task	Task name (options: cube_stacking, mug_in_drawer, drill_in_box, stick_in_bin)	✅	✅	✅	✅
data_type	Data type (options: rgbd_and_mesh, rgbd, mesh)	✅	✅	✅	✅
feature_type	Feature type (options: radio_v25_b, clip_resnet50_fpn, dino_v2_vits14, rgb)	✅	✅	✅	✅
add_external_cam	Whether to add an external camera as additional input to the model (and mapping)	✅	✅	✅	✅
demos_datagen	Demonstration range for data generation (supports ranges like 0-9)	✅	❌	❌	❌
visualize	Whether to visualize the data generation, training, or closed loop evaluation process (open loop evaluation is always visualized)	✅	✅	❌	✅
output_dir	Path to the output directory to create the dataset	✅	❌	❌	❌
hdf5_file	Path to the HDF5 file containing simulation environment and demonstration trajectories	✅	❌	❌	✅
headless	Whether to run the simulation in headless mode	✅	❌	❌	✅
base_log_dir	Path for storing Weights & Biases logs and checkpoints	❌	✅	✅	❌
dataset	Path to the dataset directory to load (in closed loop evaluation only needed in combination with --demo_mode execute_gt_goals)	❌	✅	✅	✅
demos_train	Demonstration range of the training set (supports ranges like 0-9)	❌	✅	❌	❌
demos_valset	Demonstration range of the validation set during training (supports ranges like 0-9)	❌	✅	❌	❌
train_iters	Number of training iterations	❌	✅	❌	❌
batch_size	Batch size for the training set	❌	✅	❌	❌
batch_size_val	Batch size for the validation set	❌	✅	❌	❌
demos_open_loop	Demonstration range for open loop evaluation (supports ranges like 0-9)	❌	❌	✅	❌
checkpoint	Path to the .pth checkpoint file to load (optional in case of open loop evaluation and not needed in closed loop evaluation with --demo_mode execute_gt_goals)	❌	❌	✅	✅
--ignore_model_args_json	Whether to ignore the model arguments JSON file (training_args.json). Per default, the model arguments are loaded from the file to ensure consistency between training and evaluation.	❌	❌	✅	✅
demos_closed_loop	Demonstration range for closed loop evaluation (supports ranges like 0-9)	❌	❌	❌	✅
demo_mode	How to run closed loop evaluation (options: execute_gt_goals for running GT goals from a dataset and closed_loop_wait for running inference once the last predicted goal is reached)	❌	❌	❌	✅
record_videos	Whether to record the closed loop evaluation runs (specify the output directory with --record_camera_output_path <VIDEO_OUTPUT_DIR>)	❌	❌	❌	✅
record_camera_output_path	Path to the directory to store the recorded closed loop evaluation videos (only used in combination with --record_videos)	❌	❌	❌	✅

Note

If you set parameters differently from default, you need to ensure they stay compatible across workflow steps. For example, when creating a dataset with --data_type mesh and/or --feature_type rgb, training on that dataset will only work if you also set these parameters for training.

Note

For detailed parameter definitions and their complete applicability, refer to the args.py file. Each workflow step has its own parameter class (e.g., TrainingAppArgs for training) that inherits from parent classes.