Meta-Residual Policy Learning

This repo contains code accompaning the paper, Meta-Residual Policy Learning: Zero-trial Robot Skill Adaptation via Knowledge Fusion (IEEE RA-L submission). It includes code for running the robotic peg-in-hole assembly tasks. This repository is based on PEARL.

Dependencies

We recommend using conda create environment with

conda env create -f mrplenv.yaml

This installation has been tested only on 64-bit Ubuntu 16.04.

Usage

To reproduce an experiment, run:

python launch_experiment.py ./configs/pih-meta.json

Output files will be written to ./output/pih-meta/[EXP NAME] where the experiment name is uniquely generated based on the date. The file progress.csv contains statistics logged over the course of training. To visualize learning curves, run:

python viskit/viskit/frontend.py output/pih-meta/

For evaluating the learned model, run

python sim_policy.py ./configs/pih-meta.json ./output/pih-meta/[EXP NAME] --num_trajs=20

To visualize the evaluation results, modify variable expdir=output/pih-meta/[EXP NAME]/eval_trajectories/ in plot_fig.py, and run

python plot_fig.py

---

Contact

To ask questions or report bugs, please open an issue.