Skip to main content

​
Overview

The parc_0_setup_iter.py script initializes and configures all the necessary files for running a complete PARC iteration. It creates configuration files for all four main PARC stages (train gen, kin gen, tracker, phys record) based on a master setup configuration.

​
Purpose

This script:
  • Generates configuration files for each of the 4 PARC pipeline stages
  • Sets up directory structure for outputs
  • Configures dataset paths and model checkpoints
  • Computes sampling weights for different motion types
  • Pre-computes terrain data for augmentation during MDM training

​
Usage

​
Basic Command

python scripts/parc_0_setup_iter.py --config path/to/setup_config.yaml

​
Default Configuration

If no config is provided, the script uses the default: 
python scripts/parc_0_setup_iter.py
# Uses: data/configs/parc_0_setup_iter_config.yaml

​
Command-Line Arguments

ArgumentRequiredDescription
--configNoPath to the setup configuration YAML file

​
Configuration Parameters

​
Output Control

  • output_dir: Base directory for all PARC iteration outputs
  • write_train_gen: Boolean flag to write train gen configs
  • write_kin_gen: Boolean flag to write kinematic generation configs
  • write_tracker: Boolean flag to write tracker configs
  • write_phys_record: Boolean flag to write physics recording configs

​
Stage 1: Train Gen Configuration

  • input_mdm_config_path: Path to MDM training config template
  • input_mdm_model_path: Path to previous MDM model checkpoint (null for first iteration)
  • input_sampler_stats_path: Path to sampler statistics file for normalization (null for first iteration)
  • input_create_dataset_config_path: Path to dataset creation config
  • iter_start_dataset_path: Path where the iteration dataset will be saved
  • input_dataset_folder_paths: List of folders containing motion data

​
Stage 2: Kin Gen Configuration

  • input_kin_gen_config_path: Path to kinematic generation config template
  • kin_gen_num_batches_of_motions: Number of parallel generation jobs (e.g., 10)
  • kin_gen_num_motions_per_batch: Motions generated per batch (e.g., 50)
  • kin_gen_motion_id_offset: Starting ID for generated motions (e.g., 0, 3500)
  • kin_gen_save_name: Base name for saved motion files
  • kin_gen_procgen_mode: Terrain generation mode: “BOXES”, “PATHS”, “STAIRS”, or “FILE”
  • kin_gen_start_heading_mode: Character heading initialization: “auto” or “random”

​
Stage 3: Tracker Configuration

  • input_tracker_config_path: Path to tracker training config template
  • input_tracker_model_path: Path to previous tracker model (null for first iteration)

​
Stage 4: Phys Record Configuration

  • input_phys_record_config_path: Path to physics recording config template

​
Example Configuration

output_dir: "$DATA_DIR/experiments/parc/iteration_1/"

write_train_gen: true
write_kin_gen: true
write_tracker: true
write_phys_record: true

##### PARC 1 - TRAIN GEN #####
input_mdm_config_path: "data/configs/parc_1_train_gen_default.yaml"
input_mdm_model_path: null  # First iteration
input_sampler_stats_path: null

input_create_dataset_config_path: "data/configs/create_dataset_config.yaml"
iter_start_dataset_path: "$DATA_DIR/experiments/parc/iteration_1/motions.yaml"
input_dataset_folder_paths:
  - "$DATA_DIR/initial_dataset/"

##### PARC 2 - KIN GEN #####
input_kin_gen_config_path: "data/configs/parc_2_kin_gen_default.yaml"
kin_gen_num_batches_of_motions: 10
kin_gen_num_motions_per_batch: 50
kin_gen_motion_id_offset: 0
kin_gen_save_name: "terrain_motion"
kin_gen_procgen_mode: "BOXES"
kin_gen_start_heading_mode: "random"

##### PARC 3 - TRACKER #####
input_tracker_config_path: "data/configs/parc_3_tracker_default.yaml"
input_tracker_model_path: null  # First iteration

##### PARC 4 - PHYS RECORD #####
input_phys_record_config_path: "data/configs/parc_4_phys_record_default.yaml"

​
Output Structure

The script creates the following directory structure: 
output_dir/
├── p1_train_gen/
│   ├── mdm_config.yaml
│   ├── mdm_config_<timestamp>.yaml
│   └── create_dataset_config.yaml
├── p2_kin_gen/
│   ├── <motion_name>_<start_id>_<end_id>/
│   │   └── kin_gen_config.yaml
│   └── ignore/raw/...
├── p3_tracker/
│   ├── tracker.yaml
│   ├── tracker_<timestamp>.yaml
│   └── create_dataset_config.yaml
└── p4_phys_record/
    ├── phys_record.yaml
    ├── phys_record_<timestamp>.yaml
    └── create_dataset_config.yaml

​
Usage in PARC Pipeline

This script should be run before starting each PARC iteration: 
# 1. Set up configs for iteration
python scripts/parc_0_setup_iter.py --config my_parc_setup.yaml

# 2. Run the four PARC stages
python scripts/parc_1_train_gen.py --config output_dir/p1_train_gen/mdm_config.yaml
python scripts/parc_2_kin_gen.py --config output_dir/p2_kin_gen/<motion_batch>/kin_gen_config.yaml
python scripts/parc_3_tracker.py --config output_dir/p3_tracker/tracker.yaml
python scripts/parc_4_phys_record.py --config output_dir/p4_phys_record/phys_record.yaml

​
Location

scripts/parc_0_setup_iter.py

Build docs developers (and LLMs) love

Get started for freeTalk to us