Single-Object Grasp Guess Example

This example shows how to generate collision-free grasp guesses for a single object using grasp_guess.py. It assumes you have a working gripper configuration and an object mesh.

Prerequisites

• A valid gripper USD and configuration (e.g., onrobot_rg6)
• An object mesh file in USD, OBJ, or STL format

Quick Start

python scripts/graspgen/grasp_guess.py \
    --gripper_config onrobot_rg6 \
    --object_file objects/banana.obj \
    --num_grasps 1024

Notes:

• If the gripper definition .npz does not exist yet, it will be created automatically.
• OBJ/STL inputs are used directly. USD inputs are converted to OBJ internally for grasp guessing.

Common Options

• --seed <int>: Make results reproducible.
• --num_orientations <int>: Try multiple rotations per surface point.
• --percent_random_guess_angle <float 0..1>: Mix of axis-aligned vs random rotations.
• --standoff_distance <float> and --num_offsets <int>: Control initial finger placement and retries.

Example with more control:

python scripts/graspgen/grasp_guess.py \
    --gripper_config onrobot_rg6 \
    --object_file objects/banana.obj \
    --seed 123 \
    --num_grasps 2048 \
    --num_orientations 8 \
    --percent_random_guess_angle 0.25 \
    --standoff_distance 0.0015 \
    --num_offsets 24

Output

Results are written in Isaac Grasp YAML format, typically under grasp_guess_data/<gripper>/object.yaml. Each grasp entry contains a transform and joint cspace_position/pregrasp_cspace_position values. See the component docs for details.

Visualization

• Web-based, fast: visualize_grasp_data.py (recommended)

python scripts/graspgen/tools/visualize_grasp_data.py \
    --grasp-paths grasp_guess_data/onrobot_rg6/banana.yaml \
    --object-root .

• Full USD gripper: grasp_display.py (slower with many grasps)

python scripts/graspgen/grasp_display.py \
    --grasp_file grasp_guess_data/onrobot_rg6/banana.yaml \
    --max_num_grasps 100