LearningSamplingDistributionsUsingCVAE

Requirements:

• Python 3.x (I confirm that the scripts perfectly run in Python 3.8.10)
• pytorch 1.9.1+cu111
• numpy 1.20.3
• wandb (optional): if you don't want to use wandb for tracking the training process, type 'no' in wandb at globals.ini in configs/.
• matplotlib 3.3.0 (for visualizing results)
• colorama 0.4.3
• scipy 1.7.3
• imageio & imageio--ffmpeg 2.14.1 & 0.4.5
• moviepy 1.0.3
• sklearn
• cv2

Descrition:

This repository aims to learn the sampling distribution of RRT or RRT* algorithm.
It implements the learned sampling module, which samples a random sample that may lie at the optimal path.
The optimal path is predicted by the conditional variational auto-encoder (CVAE) model.
The model gets an environment image as a conditional value and generates random samples.
With a probability _lambda, the learned sample is used. Otherwise, uniform sampling is applied.
The detailed descriptioin is on this paper.

Dataset:

• motion_planning_datasets
  : motion_planning_datasets include several different environments.

How to use:

• Gather a data

{repository} $ python{PYTHONV} learned_sampling.py --data_gather True

This step gather a data which contains environment images and the samples that may lie at the optimal path.
This (sub) optimal path is obtained with RRT* within a constant iteration.
Output results is '{environment_name}_env_sample.pickle'. The pickle file is used to train the CVAE model.

• Train a model

{repository} $ python{PYTHONV} learned_sampling.py --train True

• Test a model

{repository} $ python{PYTHONV} learned_sampling.py --test True --model_path ./checkpoints/{SOME_CHECKPOINT_FILE}.pt

Trained results

The CVAE model predict the (sub- or near-) optimal path at the current environment ('shifting_gaps').
The random configuration is sampled with 0.5 probability of uniform and 0.5 of bias by CVAE model.

futher detail of this repo.

• learned_sampling.py: main fuction for running whole scripts.
• cvae_sampler.py: Training / Evaluating / Testing a model.
• env_2d.py: a basic class for dealing with the environments, such as visualizing or collision checking, etc.
• rrt2d.py: rapidly-exploring random tree (RRT) path planning algorithm. It assumes a 2d-environment.
• rrtstar2d.py: optimal-variant RRT (RRT*) path planning algorithm. It assumes a 2d-environment.
• rrt_algorithm_example.py: Examples for using RRT algorithm.
• utils.py: some utitlity functions, mainly preprocessing data.

References.

Please contact me ([email protected]) if you have any questions.