rope-manipulation

Manipulating rope to different goal states using Mujoco.

Quick Start

1. Create conda environment + install dm_control

conda env create --file environment.yaml

Install dm_control

2. Run the tuning.py to view the PID controller tuning results

python tuning.py

4. In the mujoco folder, run explorer_dataset.py to generate the dataset

python explorer_dataset.py

For the training, make sure to comment out the wandb functions such as log, init as it won't work on machines other than the owners.

5. In the learning folder, run train_inv_dyn.py to run the training for the inverse dynamics model

python train_inv_dyn.py

5. In the learning folder, run train_infogan.py to run the training for the infogan model

python train_infogan.py

Run

To-do

• Basic env with rope + panda
• Position Controller for Panda
• Learn low-dim plan representation from image start/goal configurations according to Learning Plannable Representations with Causal InfoGAN
• Execute plans using MPC

Sources

• Panda Mujoco xml model adapted from https://github.com/justagist/mujoco_panda