Flow-based Domain Randomization

Flow-based Domain Randomization for Learning and Sequencing Robotic Skills

How to run

Set up and activate a conda environment that works with isaaclab

python -m pip install -e .

To run, simply execute the following command from the root directory

python train_rl.py --task=<env>-<method>-v0 --headless

where env in {Ant, Anymal, Cartpole, Humanoid, Gears, Quadcopter} and method in {GOFLOW, DORAEMON, LSDR, FullDR, ADR, NoDR}

For example,

python train_rl.py --task=Gears-GOFLOW-v0 --headless

Remove --headless to visualize. For the gears task, you will need to reduce the number of environments to visualize. To reproduce the experiments in the paper, see goflow/main_experiments.py

All checkpoints and tensorboard logs get saved to the logs directory. You can visualize a training checkpoint with

python train_rl.py --task=<env>-<method>-v0 --checkpoint=<path-to-checkpoint>

Cite this paper

@inproceedings{curtis2025flowbaseddomainrandomizationlearning,
  title     = {Flow-based Domain Randomization for Learning and Sequencing Robotic Skills},
  author    = {Aidan Curtis and Eric Li and Michael Noseworthy and Nishad Gothoskar and Sachin Chitta and Hui Li and Leslie Pack Kaelbling and Nicole Carey},
  booktitle = {Proceedings of the 41st International Conference on Machine Learning (ICML)},
  year      = {2025},
  note      = {To appear},
  url       = {https://arxiv.org/abs/2502.01800},
  archivePrefix = {arXiv},
  eprint    = {2502.01800},
  primaryClass = {cs.RO}
}