Puzzle Generator

Generate synthetic datasets for robot manipulation tasks using procedural content generation (PCG). The generator engine is written in Python and utilizes Blender's scripting API together with the Phobos add-on.

Generated Puzzles:

• use PCG
• are manipulation challenges
• are generated according to a given config with constraints
• include stochasticity within those constraints
• have interlocking dependencies between joints
• can be used for training or benchmarking an agent/algorithm
• are URDF files but the output can be anything that is supported by Blender/Phobos

Here is an example of a generated grid world puzzle being solved by a fetch robot. The translucent green box marks the goal position for the green link:

Some of the sampling algorithms need PyBullet with OMPL for evaluation during the sampling process. The others can be used without the installation of PyBullet and OMPL

Install (limited)

• Download and install Blender 2.93
• Install the Phobos add-on for Blender from the correct branch for version 2.93

Features: create_custom_urdf.py, SimpleSlidersSampler, GridWorldSampler, Lockbox2017Sampler, EscapeRoomSampler, MoveTwiceSampler, MoveNTimesSampler, RoomsSampler

Install (complete)

The above steps plus:

• Install OMPL
• Install PyBullet

Features: pybullet_simulation.py, ContinuousSpaceSampler, LockboxRandomSampler

Run

There are two possible scripts that you can run. You can either create a custom URDF file or use the puzzle generator to generate a puzzle:

Create a custom URDF

Adjust the script create_custom_urdf.py to your requirements and run it with:

./blender-2.93 --background --python create_custom_urdf.py

Create a puzzle

Adjust the script puzzle_generator.py to your requirements and run it with:

./blender-2.93 --background --python puzzle_generator.py