- 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
- From the project root:
pip install -e .
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.
- You need a mujoco license file. Move the text license into the root
fetch-block-construction
directory, and rename it asmjkey.txt
- From the project root, run:
docker build -t <docker_username>/<image_name>:<image_tag> .
docker push <docker_username>/<image_name>:<image_tag>
Initial stack environment code based on [gym-fetch-stack](https://github.com/CDMCH/gym-fetch-stack
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}
}