Constrained Action-Space Representation for Physical Manipulation

This is the repository for Saumya Shah and Matthew Goodman's ESE 650 Final Project (2020).

Credit to Lingzhi Zhang and Andong Cao for sharing their code.

@misc{zhang2019neural,
    title={Neural Embedding for Physical Manipulations},
    author={Lingzhi Zhang and Andong Cao and Rui Li and Jianbo Shi},
    year={2019},
    eprint={1907.06143},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

Credit to Tianhong Dai for his implementation of HER in OpenAI Gym robotics environments. https://github.com/TianhongDai/hindsight-experience-replay

Getting Started

Pretrained HER

Download Tianhong Dai's pretrained HER models from: https://drive.google.com/open?id=18MEEwweR8Ad1t1yxgTwI1VYmIsVPttXk

mkdir models/her_pretrained/FetchPush-v1
mv model.pt models/her_pretrained/FetchPush-v1

Generating trajectories

Argument documentation found with python generate_trajectories.py -h

To generate trajectories with default arguments (1 file, 1000 trajectories for file, saved to /data):

python generate_trajectories.py

Semi-random forward goal trajectories. Object is placed, goal then placed in pi/3 radius in front of object.

python generate_trajectories.py --trajectory-length 8 --num_trajectory_per_file 1000 --filename_start_idx 1 --num_files 20 --outdir data/forward_inline/train --image-shape 128 128 --simplify-task

Inline goal trajectories:

python generate_trajectories.py --trajectory-length 8 --num_trajectory_per_file 1000 --filename_start_idx 1 --num_files 20 --outdir data/forward_inline/train --image-shape 128 128 --simplify-task --goal-inline

View individual trajectory

To view an individual trajectory:

python -m utils.trajectory_loader _path_to_data_

Training the forward model

python train_forward_model.py --trajectory-length 8 --train-data-path forward_inline_data/train --gpu-id 2 --log-port 8081 --forward-save-path models/inline_task_forward

Training the GAN

python train_gan.py --trajectory-length 8 --train-data-path forward_inline_data/train --evaluation-data-path forward_inline_data/eval --gpu-id 2 --log-port 8081 --gan-save-path models/inline_task_gan

Project Report :

https://drive.google.com/file/d/1QJucw3DEFHWvWO7DGr6RDaVVHSt_Dl6e/view