README.md

# XIRL

[![python](https://img.shields.io/badge/python-3.8-blue.svg)](https://www.python.org/downloads/release/python-383/)
[![arXiv](https://img.shields.io/badge/arXiv-2106.03911-b31b1b.svg)](https://arxiv.org/abs/2106.03911)
[![License](https://img.shields.io/badge/License-Apache%202.0-green.svg)](https://github.com/google-research/google-research/blob/master/LICENSE)

- [Overview](#overview)
- [Setup](#setup)
- [Datasets](#datasets)
- [Code Navigation](#code-navigation)
- [Experiments: Reproducing Paper Results](#experiments-reproducing-paper-results)
- [Extending XIRL](#extending-xirl)
- [Acknowledgments](#acknowledgments)

<p align="center">
  <img src="./images/teaser.gif" width="75%"/>
</p>

## Overview

Code release for our CoRL 2021 conference paper:

<table><tr><td>
    <strong>
        <a href="https://x-irl.github.io/">
            XIRL: Cross-embodiment Inverse Reinforcement Learning
        </a><br/>
    </strong>
    Kevin Zakka<sup>1,3</sup>, Andy Zeng<sup>1</sup>, Pete Florence<sup>1</sup>, Jonathan Tompson<sup>1</sup>, Jeannette Bohg<sup>2</sup>, and Debidatta Dwibedi<sup>1</sup><br/>
    Conference on Robot Learning (CoRL) 2021
</td></tr></table>

<sup>1</sup><em>Robotics at Google,</em>
<sup>2</sup><em>Stanford University,</em>
<sup>3</sup><em>UC Berkeley</em>

---

This repository serves as a general-purpose library for (a) **self-supervised pretraining** on video data and **(b)** downstream **reinforcement learning** using the learned representations as reward functions. It also contains models, training scripts and config files for reproducing our results and as a reference for implementation details.

Our hope is that the code's modularity allows you to easily extend and build on top of our work. To aid in this effort, we're releasing two additional standalone libraries:

* [x-magical](https://github.com/kevinzakka/x-magical): our Gym-like benchmark extension of MAGICAL geared towards cross-embodiment imitation.
* [torchkit](https://github.com/kevinzakka/torchkit): a lightweight library containing useful PyTorch boilerplate utilities like logging and model checkpointing.

For the latest updates, see: [x-irl.github.io](https://x-irl.github.io)

## Setup

We use Python 3.8 and [Miniconda](https://docs.conda.io/en/latest/miniconda.html) for development. To create an environment and install dependencies, run the following steps:

```bash
# Clone and cd into xirl.
git clone git@github.com:google-research/google-research.git --depth=1
cd google-research/xirl

# Create and activate environment.
conda create -n xirl python=3.8
conda activate xirl

# Install dependencies.
pip install -r requirements.txt
```

## Datasets

**X-MAGICAL**

Run the following bash script to download the demonstration dataset for the X-MAGICAL benchmark:

```bash
bash scripts/download_xmagical_dataset.sh
```

The dataset will be located in `/tmp/xirl/datasets/xmagical`. You are free to modify the save destination, just make sure you update `config.data.root` in the pretraining config file &mdash; see `base_configs/pretrain.py`.

**X-REAL**

Our real-world dataset X-REAL will be released as soon as it gets approval, stay tuned!

## Code Navigation

At a high-level, our code relies on two important but generic python scripts: `pretrain.py` for pretraining and `train_policy.py` for reinforcement learning. We use [ml_collections](https://github.com/google/ml_collections) to parameterize these scripts with experiment-specific config files. **All experiments must use config files that inherit from the base config files in `base_configs/`**. Specifically, pretraining experiments must inherit from `base_configs/pretrain.py` and RL experiments must inherit from `base_configs/rl.py`.

The rest of the codebase is organized as follows:

* `configs/` contains all config files used in our CoRL submission. They inherit from `base_configs/`.
* `xirl/` is the core pretraining codebase.
* `sac/` is the core Soft-Actor-Critic implementation adapted from [pytorch_sac](https://github.com/denisyarats/pytorch_sac).
* `scripts/` contains miscellaneous bash scripts.

## Experiments: Reproducing Paper Results

**Core Scripts**

- [x] Same-embodiment setting (Section 5.1)
    - [x] Pretraining: `python pretrain_xmagical_same_embodiment.py --help`
    - [x] RL: `python rl_xmagical_learned_reward.py --help`
- [x] Cross-embodiment setting (Section 5.2)
    - [x] Pretraining: `python pretrain_xmagical_cross_embodiment.py --help`
    - [x] RL: `python rl_xmagical_learned_reward.py --help`
- [x] RL with environment reward
    - [x] `python rl_xmagical_env_reward.py --help`
- [x] Interactive reward visualization (Section 5.4)
    - [x] `python interact_reward.py --help`

**Misc. Scripts**

- [x] Visualize dataloader for frame sampler debugging
    - [x] `python debug_dataset.py --help`
- [x] Compute goal embedding with a pretrained model
    - [x] `python compute_goal_embedding.py --help`
- [x] Quick n' dirty multi-GPU RL training
    - [x] With environment reward: `bash scripts/launch_rl_multi_gpu.sh`

## Extending XIRL

> How can I implement my own self-supervised pretraining algorithm?

You'll want to inherit from `xirl.trainers.base.Trainer` and implement the `__init__` and `compute_loss` methods. For reference, take a look at `xirl/trainers/tcc.py` to see how Temporal Cycle Consistency is implemented. Make sure to add your new algorithm to the `TRAINERS` dict in `factory.py`.

> How do I modify the way frames are sampled in the dataloader?

Create your own sampler in `xirl/frame_samplers.py` and add it to the `FRAME_SAMPLERS` dict in `factory.py`.

> How can I implement additional pretraining evaluation metrics?

You'll want to inherit from `xirl.evaluators.base.Evaluator` class and as you guessed, add it to the `EVALUATORS` dict in `factory.py`. See `xirl/evaluators` for our current list of qualitative and quantative evaluation metrics

## Acknowledgments

Many people have contibuted one way or another in the making and shaping of this repository. In no particular order, we'd like to thank [Alex Nichol](https://aqnichol.com/), [Nick Hynes](https://www.linkedin.com/in/nhynes-), [Brent Yi](https://brentyi.com/), [Jimmy Wu](https://www.cs.princeton.edu/~jw60/) and [Sam Toyer](https://scholar.google.com.au/citations?user=J8E8GQYAAAAJ&hl=en) for their fruitful back-and-forth discussions.

## Citation

If you find this code useful, consider citing our work:

```bibtex
@inproceedings{zakka2021xirl,
  author    = {Zakka, Kevin and Zeng, Andy and Florence, Pete and Tompson, Jonathan and Bohg, Jeannette and Dwibedi, Debidatta},
  title     = {XIRL: Cross-embodiment Inverse Reinforcement Learning},
  booktitle = {Proceedings of the 5th Conference on Robot Learning (CoRL)},
  year      = {2021},
}
```