Skip to main content

​
Overview

The parc_2_kin_gen.py script generates new terrain-traversal motions using the trained Motion Diffusion Model. It creates procedural terrains, plans paths, generates candidate motions, applies kinematic selection heuristics, and optimizes the results.

​
Purpose

This is Stage 2 of the PARC pipeline. It:
  • Generates procedural terrains (boxes, paths, stairs, or from files)
  • Plans traversal paths using A* pathfinding
  • Generates motion candidates autoregressively along paths
  • Filters motions based on contact and penetration losses
  • Applies kinematic optimization to refine motions
  • Removes hesitation frames
  • Saves both original and mirrored versions of motions

​
Usage

​
Basic Command

python scripts/parc_2_kin_gen.py --config path/to/kin_gen_config.yaml

​
Default Configuration

python scripts/parc_2_kin_gen.py
# Uses: data/configs/parc_2_kin_gen_default.yaml

​
Command-Line Arguments

ArgumentRequiredDescription
--configNoPath to the kinematic generation configuration YAML file

​
Key Configuration Parameters

​
Generation Settings

  • mdm_model_path: Path to trained MDM model checkpoint directory
  • num_new_motions: Total number of motions to generate
  • motion_id_offset: Starting ID for numbering generated motions
  • save_name: Base name for saved motion files
  • output_dir: Directory for saving raw generated motions

​
Procedural Terrain Settings

  • procgen_mode: Terrain generation mode: “BOXES”, “PATHS”, “STAIRS”, or “FILE”
  • new_terrain_dim_x: Terrain width in grid cells
  • new_terrain_dim_y: Terrain length in grid cells
  • dx: Grid cell size in x-direction (meters)
  • dy: Grid cell size in y-direction (meters)
  • simplify_terrain: Boolean flag to apply terrain simplification
  • first_heading_mode: Character start heading: “auto” or “random”

​
Terrain Type Configurations

​
BOXES Mode

boxes:
  num_boxes: 10
  min_box_h: -3.0
  max_box_h: 3.0
  box_max_len: 10
  box_min_len: 5
  max_box_angle: 6.28318530718  # 2π
  min_box_angle: 0.0

​
PATHS Mode

paths:
  num_terrain_paths: 4
  maxpool_size: 1
  path_min_height: -2.8
  path_max_height: 3.0
  floor_height: -3.0

​
STAIRS Mode

stairs:
  min_stair_start_height: -3.0
  max_stair_start_height: 1.0
  min_step_height: 0.15
  max_step_height: 0.25
  num_stairs: 4
  min_stair_thickness: 2.0
  max_stair_thickness: 8.0

​
FILE Mode

  • input_terrain_path: Path to pickle or YAML file containing pre-made terrains

​
Motion Quality Filters

  • max_contact_loss: Maximum allowed contact loss for valid motions
  • max_pen_loss: Maximum allowed penetration loss
  • max_total_loss: Maximum allowed total loss

​
Path Planning (A* Settings)

astar:
  min_start_end_xy_dist: 20.0  # Minimum distance between start and end
  # Additional A* parameters...

​
MDM Path Generation Settings

mdm_path:
  mdm_batch_size: 32  # Number of candidate motions per generation
  top_k: 3  # Number of best motions to keep
  # Additional generation parameters...

​
Optimization Settings

  • use_opt: Boolean flag to enable kinematic optimization
  • remove_hesitation: Boolean flag to remove hesitation frames
  • only_gen: Boolean flag to only generate terrains without motions (for debugging)

​
Optimization Parameters (when use_opt: true)

opt:
  device: "cuda:0"
  output_dir: "path/to/optimized_output"
  char_model: "path/to/character_model.yaml"
  num_iters: 100
  step_size: 0.01
  
  # Loss weights
  w_root_pos: 1.0
  w_root_rot: 1.0
  w_joint_rot: 1.0
  w_smoothness: 0.1
  w_penetration: 10.0
  w_contact: 5.0
  w_sliding: 2.0
  w_body_constraints: 1.0
  w_jerk: 0.5
  max_jerk: 100.0
  
  use_wandb: false
  auto_compute_body_constraints: true
  
  # Character point sampling for collision detection
  char_point_samples:
    sphere_num_subdivisions: 2
    box_num_slices: 3
    box_dim_x: 0.5
    box_dim_y: 0.5
    capsule_num_circle_points: 8
    capsule_num_sphere_subdivisions: 2
    capsule_num_cylinder_slices: 3

​
Generation Pipeline

For each motion to generate:
  1. Generate Terrain: Create procedural terrain based on procgen_mode
  2. Plan Path: Use A* to find traversal path on terrain
  3. Generate Candidates: Use MDM to generate mdm_batch_size candidate motions
  4. Filter by Quality: Keep motions below loss thresholds
  5. Select Top K: Choose the top_k best motions
  6. Optimize: Apply kinematic optimization (if enabled)
  7. Post-process: Remove hesitation frames (if enabled)
  8. Save: Save optimized motion and flipped version with terrain data

​
Example Configuration

# Model and generation
mdm_model_path: "$DATA_DIR/iteration_1/p1_train_gen/checkpoints"
num_new_motions: 50
motion_id_offset: 0
save_name: "terrain_motion"
output_dir: "$DATA_DIR/iteration_1/p2_kin_gen/raw"

# Terrain generation
procgen_mode: "BOXES"
new_terrain_dim_x: 100
new_terrain_dim_y: 100
dx: 0.05
dy: 0.05
simplify_terrain: true
first_heading_mode: "random"

# Quality thresholds
max_contact_loss: 0.5
max_pen_loss: 0.3
max_total_loss: 1.0

# Optimization
use_opt: true
remove_hesitation: true
only_gen: false

# A* pathfinding
astar:
  min_start_end_xy_dist: 20.0

# MDM generation
mdm_path:
  mdm_batch_size: 32
  top_k: 3

mdm_gen:
  # MDM generation settings...

# Terrain type configs
boxes:
  num_boxes: 10
  min_box_h: -3.0
  max_box_h: 3.0
  box_max_len: 10
  box_min_len: 5

# Optimization settings
opt:
  device: "cuda:0"
  output_dir: "$DATA_DIR/iteration_1/p2_kin_gen/optimized"
  num_iters: 100
  use_wandb: false
  # ... additional opt params

​
Output Files

For each generated motion batch: 
output_dir/
├── <motion_name>_<id>_0.pkl
├── <motion_name>_<id>_1.pkl
├── <motion_name>_<id>_2.pkl  # top_k variants
└── ...

opt/output_dir/  # If optimization enabled
├── <motion_name>_<id>_0_opt.pkl
├── <motion_name>_<id>_1_opt.pkl
├── <motion_name>_<id>_2_opt.pkl
├── flipped/
│   ├── <motion_name>_<id>_0_opt_flipped.pkl
│   ├── <motion_name>_<id>_1_opt_flipped.pkl
│   └── ...
└── log/
    ├── log_<motion_name>_<id>_0.txt
    └── ...
Each .pkl file contains:
  • Motion frames (root position, rotation, joint rotations, contacts)
  • Terrain heightfield data
  • Path nodes
  • Heightfield mask indices
  • Loss values and metadata

​
Parallel Batch Execution

The parc_0_setup_iter.py script generates multiple config files for parallel execution: 
# Run multiple batches in parallel (e.g., on a cluster)
python scripts/parc_2_kin_gen.py --config output/p2_kin_gen/batch_0_49/kin_gen_config.yaml &
python scripts/parc_2_kin_gen.py --config output/p2_kin_gen/batch_50_99/kin_gen_config.yaml &
python scripts/parc_2_kin_gen.py --config output/p2_kin_gen/batch_100_149/kin_gen_config.yaml &

​
Implementation Details

​
Key Modules

  • parc/motion_synthesis/procgen/astar.py: A* pathfinding on terrain graphs
  • parc/motion_synthesis/procgen/mdm_path.py: Autoregressive motion generation along paths
  • parc/motion_synthesis/motion_opt/motion_optimization.py: Kinematic optimization
  • parc/util/terrain_util.py: Terrain generation and manipulation
  • parc/util/motion_edit_lib.py: Motion editing utilities (hesitation removal, flipping)

​
Usage in PARC Pipeline

# Stage 1: Train generator
python scripts/parc_1_train_gen.py --config configs/train_gen.yaml

# Stage 2: Generate motions (THIS SCRIPT)
python scripts/parc_2_kin_gen.py --config configs/kin_gen.yaml

# Stage 3: Train tracker
python scripts/parc_3_tracker.py --config configs/tracker.yaml

​
Location

scripts/parc_2_kin_gen.py

Build docs developers (and LLMs) love

Get started for freeTalk to us