Virtual Creatures Re-Implementation

Installing

This repo depends on a few popular packages:

• PyBullet (for physics simulation)
• PyTorch (for creature controller)
• NumPy (dealing w/ positions, math, etc.)
• odio_urdf (for creating and manipulating URDF XML)

Running

This repo contains two main scripts, one for evolving virtual creatures and one for testing evolved creatures. To train the creatures run:

$ python train.py --help

usage: train.py [-h] [--task TASK] [--workers WORKERS] [--pop_size POP_SIZE] [--top_k TOP_K] [--mutate MUTATE] [--num_gens NUM_GENS] [--out_dir OUT_DIR]

evolve virtual creatures

optional arguments:
  -h, --help           show this help message and exit
  --task TASK          walking or following
  --workers WORKERS    number of workers in multi-processing pool
  --pop_size POP_SIZE  number of creatures in the population
  --top_k TOP_K        number of creatures used for mutation
  --mutate MUTATE      number of mutated creatures to generate
  --num_gens NUM_GENS  number of generations to run
  --out_dir OUT_DIR    directory to save evolved creature files

To run the walking task with default settings simply run python train.py. After running this script the best creature will be logged (by default to the models/ directory). The logged creatures are python objects stored as pickle files. These files can then be tested using the test script:

$ python test.py -h

usage: test.py [-h] [--creature CREATURE] [--task TASK] [--record] [--out_path OUT_PATH]

test a generated creature

optional arguments:
  -h, --help           show this help message and exit
  --creature CREATURE  path of stored creature pickle file
  --task TASK          walking or following
  --record             wether or not to record the simulation
  --out_path OUT_PATH  name of .mp4 file for recording output

The evolved walking creatures can be ran using the following command:

$ python test.py --creature models/best_100.pkl

Some examples of evolved creatures are included in the examples/ directory