richardrl / fetch-block-construction

Environment codebase for ICRA 2020 paper "Towards Practical Multi-object Manipulation using Relational Reinforcement Learning"

Fetch Block Construction

• Produce up to 25 blocks, which is the max number supported by Mujoco 1.5
  • Max 17 blocks with uniform sampling of collision-free init positions on the table
• Set novel goal configurations with a goal position vector for each block
• Preset goal configurations
  • Single tower
  • Multiple towers
  • Pyramid
• Randomize block lengths to produce cuboids
• Optional rotational control for action space

Installation

1. From the project root: pip install -e .

Docker Build

Building the Docker image is optional, but makes it much easier to continuously deploy to EC2, GCP, etc. The Docker image is designed to support rlkit-relational. The example scripts there will ask you to input the name of your Docker image.

1. You need a mujoco license file. Move the text license into the root fetch-block-construction directory, and rename it as mjkey.txt
2. From the project root, run: docker build -t <docker_username>/<image_name>:<image_tag> .
3. docker push <docker_username>/<image_name>:<image_tag>

Credits

Initial stack environment code based on [gym-fetch-stack](https://github.com/CDMCH/gym-fetch-stack

Citation

If you find this code useful, please cite:

@inproceedings{li19relationalrl,
  Author = {Li, Richard and
  Jabri, Allan and Darrell, Trevor and Agrawal, Pulkit},
  Title = {Towards Practical Multi-object Manipulation using Relational Reinforcement Learning},
  Booktitle = {ICRA},
  Year = {2020}
}