datagen.md

Batch Data Generation

The Batch Data Generation workflow example, datagen.py, processes multiple objects to create training datasets for machine learning systems. This workflow is designed for large-scale data generation and can be customized for specific training requirements.

Overview

The batch generation workflow:

1. Processes multiple objects from a JSON configuration file
2. Applies consistent settings across all objects
3. Generates grasp candidates for each object
4. Validates grasps through physics simulation
5. Creates organized datasets with success/failure labels

Usage

Basic Batch Processing

To generate grasps for a set of objects, you need the base folder where you keep the objects (--object_root) and a JSON file with scales and object file names relative to that root.

JSON Configuration Format

The object_scales_json file should contain object paths (relative to object_root) and their scales:

{
    "objects/banana.obj": 0.75,
    "objects/apple.obj": 0.8,
    "objects/orange.usd": 1.2,
    "objects/peach.stl": 0.5
}

Directory Structure

object_root/
├── objects/
│   ├── banana.obj
│   ├── apple.obj
│   ├── orange.usd
│   └── peach.stl

Standard Dataset Generation Example

python scripts/graspgen/datagen.py \
    --gripper_config onrobot_rg6 \
    --object_scales_json objects/datagen_example.json \
    --object_root objects \
    --num_grasps 1024 

This would create the Isaac Grasp data files:

datagen_sim_data/
└── onrobot_rg6/
    ├── banana.0.75.yaml          # Banana object at 0.75 scale
    ├── banana.1.0.yaml           # Banana object at 1.0 scale
    ├── Sphere.0.05.yaml          # Sphere object at 0.05 scale
    └── threelayer.0.005.yaml     # Threelayer object at 0.005 scale

Multi-Gripper Dataset

# Generate dataset for multiple grippers
gripper_configs=("robotiq_2f_85" "onrobot_rg6" "franka_panda")
for gripper_config in "${gripper_configs[@]}"; do
    python scripts/graspgen/datagen.py \
        --gripper_config "$gripper_config" \
        --object_scales_json objects/datagen_example.json \
        --object_root objects \
        --num_grasps 1024 
done

Custom Output Directories

python scripts/graspgen/datagen.py \
    --gripper_config onrobot_rg6 \
    --object_scales_json objects/datagen_example.json \
    --object_root objects \
    --sim_output_folder custom_sim_data \
    --guess_output_folder custom_guess_data \
    --num_grasps 1024 \
    --overwrite_existing