f-IRL: Inverse Reinforcement Learning via State Marginal Matching

Appear in Conference on Robot Learning (CoRL) 2020. This repository is to reproduce the results for our method and baselines showed in the paper. arXiv link, Website link (with presentation videos)

Authors: Tianwei Ni, Harshit Sikchi, Yufei Wang, Tejus Gupta, Lisa Lee°, Benjamin Eysenbach°. where * indicates equal contribution (order by dice rolling) and ° equal advising.

Note: This repository also contains implementation for prior imitation learning methods like f-MAX, AIRL, MaxEntIRL, GAIL+SAC and Behavior Cloning.

Installation

• PyTorch 1.5+
• OpenAI Gym
• MuJoCo
• pip install ruamel.yaml
• Download expert data that are used in our paper from Google Drive as expert_data/ folder
  • states/: expert state trajectories for each environment. We obtain two sets of state trajectories for our method/MaxEntIRL/f-MAX (*.pt) and AIRL (*_airl.pt), respectively.
  • actions/: expert action trajectories for each environment for AIRL (*_airl.pt)
  • meta/: meta information including expert reward curves through training
  • reward_models/: the reward models saved from our algorithm

File Structure

• f-IRL (our method): firl/
• Baselines (f-MAX, AIRL, MaxEntIRL, GAIL+SAC, BC): baselines/
• SAC agent: common/
• Environments: envs/
• Configurations: configs/

Instructions

• All the experiments are to be run under the root folder.
• Before starting experiments, please export PYTHONPATH=${PWD}:$PYTHONPATH for env variable.
• We use yaml files in configs/ for experimental configurations, please change obj value (in the first line) for each method, here is the list of obj values:
  • Our methods (f-IRL): FKL: fkl, RKL: rkl, JS: js
  • Baselines: MaxEntIRL: maxentirl, f-MAX-RKL: f-max-rkl, GAIL: gail, AIRL: airl, BC: bc
• Please keep all the other values in yaml files unchanged to reproduce the results in our paper.
• After running, you will see the training logs in logs/ folder.

Experiments

All the commands below are also provided in run.sh.

Sec 5.1 Density task (Reacher2d)

```bash

our method and maxentirl. you can vary obj in {fkl, rkl, js, maxentirl}

python firl/irldensity.py configs/density/reachertracegauss.yml # Gaussian goal python firl/irldensity.py configs/density/reachertracemix.yml # Mixture of Gaussians goal

f-MAX-RKL, GAIL. you can vary obj in {f-max-rkl, gail}

python baselines/maindensity.py configs/density/reachertracegauss.yml # Gaussian goal python baselines/maindensity.py configs/density/reachertracemix.yml # Mixture of Gaussians goal ```

Sec 5.2 IRL benchmark (MuJoCo)

First, make sure that you have downloaded expert data into expert_data/. Otherwise, you can generate expert data by training expert policy: bash python common/train_expert.py configs/samples/experts/{env}.yml # env is in {hopper, walker2d, halfcheetah, ant}

Then train our method or baseline with provided expert data method (Policy Performance). Note that you can change the value of irl: expert_episodes: into {1, 4, 16} to reproduce the results of {1, 4, 16} trajectories setting shown in Table 3.

```bash

our method and maxentirl. you can vary obj in {fkl, rkl, js, maxentirl}

python firl/irl_samples.py configs/samples/agents/{env}.yml

baselines

python baselines/bc.py configs/samples/agents/{env}.yml # bc. set obj to bc python baselines/mainsamples.py configs/samples/agents/{env}.yml # f-max-rkl. set obj to f-max-rkl python baselines/mainsamples.py configs/samples/agents/airl/{env}.yml # airl. ```

After the training is done, you can choose one of the saved reward model to train a policy from scratch (Recovering the Stationary Reward Function). We provide a learned reward model in expert_data/reward_models/halfcheetah/ for demonstration purposes. bash python common/train_optimal.py configs/samples/experts/halfcheetah.yml

Sec 5.3.1 Downstream task

First, run f-IRL or the baselines on the pointmass gridworld with a uniform expert density: ```bash

you can change the obj in grid_uniform.yml to be {fkl, rkl, js, maxentirl}

python firl/irldensity.py configs/density/griduniform.yml

you can change the obj in grid_uniform.yml to be {f-max-rkl, gail}

python baselines/maindensity.py configs/density/griduniform.yml `` Then, the discriminator or the reward model should be saved in logs/ContinuousVecGridEnv-v0/{month}-{date}-uniform/{obj}/{detailed-time-stamp}/model/rewardmodel*.pkl`

Then update the path to the stored reward model in firl/priorreward/main.py at line 132-134, and run ```bash python firl/priorreward/main.py ``` to test the learned reward on the hard-to-explore task.

The information of the learned sac agent will be saved to data/prior_reward/potential/{save_name}_{alpha}_{prior_reward_weight}_sac_test_rets.npy

After obtain the learning results from multiple learned rewards/discriminators, firl/prior_reward/plot_image.py and firl/prior_reward/plot_reward.py can be used to create figure 4 in the paper.

Sec 5.3.2 Transfer task

First, make sure that you have downloaded expert data into expert_data/. Otherwise, you can generate expert data by training expert policy: Make sure that the env_name parameter in configs/samples/experts/ant_transfer.yml is set to CustomAnt-v0 bash python common/train_expert.py configs/samples/experts/ant_transfer.yml

Then train our method or baseline with provided expert data method (Policy Performance). python firl/irl_samples.py configs/samples/agents/ant_transfer.yml After the training is done, you can choose one of the saved reward model to train a policy from scratch (Recovering the Stationary Reward Function).

Transferring the reward to disabled Ant: We provide a learned reward model in expert_data/reward_models/ant_transfer/ for demonstration purposes. Make sure that the env_name parameter in configs/samples/experts/ant_transfer.yml is set to DisabledAnt-v0 bash python common/train_optimal.py configs/samples/experts/ant_transfer.yml

Citation and References

If you find our paper useful to your research, please cite the paper: @inproceedings{firl2020corl, title={f-IRL: Inverse Reinforcement Learning via State Marginal Matching}, author={Ni, Tianwei and Sikchi, Harshit and Wang, Yufei and Gupta, Tejus and Lee, Lisa and Eysenbach, Ben}, booktitle={Conference on Robot Learning}, year={2020} }

Parts of the codes are used from the references mentioned below:

• AIRL in part of envs/
• f-MAX in part of baselines/
• SAC in part of common/sac
• NPEET in part of utils/it_estimator.py