Self-Supervised Reward Regression (SSRR)

Codebase for CoRL 2021 paper "Learning from Suboptimal Demonstration via Self-Supervised Reward Regression " Authors: Letian "Zac" Chen, Rohan Paleja, Matthew Gombolay

Usage

Quick overview

The pipeline of SSRR includes

1. Initial IRL: Noisy-AIRL or AIRL.
2. Noisy Dataset Generation: use initial policy learned in step 1 to generate trajectories with different noise levels and criticize trajectories with initial reward.
3. Sigmoid Fitting: fit a sigmoid function for the noise-performance relationship using the data obtained in step 2.
4. Reward Learning: learn a reward function by regressing to the sigmoid relationship obtained in step 3.
5. Policy Learning: learn a policy by optimizing the reward learned in step 4.

I know this is a long README, but please make sure you read the entirety before trying out our code. Trust me, that will save your time!

Dependencies and Environment Preparations

Code is tested with Python 3.6 with Anaconda.

Required packages:

pip install scipy path.py joblib==0.12.3 flask h5py matplotlib scikit-learn pandas pillow pyprind tqdm nose2 mujoco-py cached_property cloudpickle git+https://github.com/Theano/[email protected]#egg=Theano git+https://github.com/neocxi/[email protected]#egg=Lasagne plotly==2.0.0 gym[all]==0.14.0 progressbar2 tensorflow-gpu==1.15 imgcat

Test sets of trajectories could be downloaded at Google Drive because Github could not hold files that are larger than 100MB! After downloading, please put full_demos/ under demos/.

If you are directly running python scripts, you will need to add the project root and the rllab_archive folder into your PYTHONPATH:

export PYTHONPATH=/path/to/this/repo/:/path/to/this/repo/rllab_archive/

If you are using the bash scripts provided (for example, noisy_airl_ssrr_drex_comparison_halfcheetah.sh), make sure to replace the first line to be

export PYTHONPATH=/path/to/this/repo/:/path/to/this/repo/rllab_archive/

Initial IRL

We provide code for AIRL and Noisy-AIRL implementation.

Running

Examples of running command would be

python script_experiment/halfcheetah_airl.py --output_dir=./data/halfcheetah_airl_test_1
python script_experiment/hopper_noisy_airl.py --output_dir=./data/hopper_noisy_airl_test_1 --noisy

Please note for Noisy-AIRL, you have to include the --noisy flag to make it actually sample trajectories with noise, otherwise it only changes the loss function according to Equation 6 in the paper.

Results

The result will be available in the output dir specified, and we recommend using rllab viskit to visualize it.

We also provide our run results available in data/{halfcheetah/hopper/ant}_{airl/noisy_airl}_test_1 if you want to skip this step!

Code Structure

The AIRL and Noisy-AIRL codes reside in inverse_rl/ with rllab dependencies in rllab_archive. The AIRL code is adjusted from the original AIRL codebase https://github.com/justinjfu/inverse_rl. The rllab archive was adjusted from the original rllab codebase https://github.com/rll/rllab.

Noisy Dataset Generation & Sigmoid Fitting

We implemented noisy dataset generation and sigmoid fitting together in code.

Running

Examples of running command would be

python script_experiment/noisy_dataset.py \
   --log_dir=./results/halfcheetah/temp/noisy_dataset/ \
   --env_id=HalfCheetah-v3 \
   --bc_agent=./results/halfcheetah/temp/bc/model.ckpt \
   --demo_trajs=./demos/suboptimal_demos/ant/dataset.pkl \
   --airl_path=./data/halfcheetah_airl_test_1/itr_999.pkl \
   --airl \
   --seed="${loop}"

Note that flag --airl determines whether we utilize the --airl_path or --bc_agent policy to generate the trajectory. Therefore, --bc_agent is optional when --airl present. For behavior cloning policy, please refer to https://github.com/dsbrown1331/CoRL2019-DREX.

The --airl_path always provide the initial reward to criticize the generated trajectories no matter whether --airl present.

Results

The result will be available in the log dir specified.

We also provide our run results available in results/{halfcheetah/hopper/ant}/{airl/noisy_airl}_data_ssrr_{1/2/3/4/5}/noisy_dataset/ if you want to skip this step!

Code Structure

Noisy dataset generation and Sigmoid fitting are implemented in script_experiment/noisy_dataset.py.

Reward Learning

We provide SSRR and D-REX implementation.

Running

Examples of running command would be

  python script_experiment/drex.py \
   --log_dir=./results/halfcheetah/temp/drex \
   --env_id=HalfCheetah-v3 \
   --bc_trajs=./demos/suboptimal_demos/halfcheetah/dataset.pkl \
   --unseen_trajs=./demos/full_demos/halfcheetah/unseen_trajs.pkl \
   --noise_injected_trajs=./results/halfcheetah/temp/noisy_dataset/prebuilt.pkl \
   --seed="${loop}"
  python script_experiment/ssrr.py \
   --log_dir=./results/halfcheetah/temp/ssrr \
   --env_id=HalfCheetah-v3 \
   --mode=train_reward \
   --noise_injected_trajs=./results/halfcheetah/temp/noisy_dataset/prebuilt.pkl \
   --bc_trajs=demos/suboptimal_demos/halfcheetah/dataset.pkl \
   --unseen_trajs=demos/full_demos/halfcheetah/unseen_trajs.pkl \
   --min_steps=50 --max_steps=500 --l2_reg=0.1 \
   --sigmoid_params_path=./results/halfcheetah/temp/noisy_dataset/fitted_sigmoid_param.pkl \
   --seed="${loop}"

The bash script also helps combining running of noisy dataset generation, sigmoid fitting, and reward learning, and repeats several times:

./airl_ssrr_drex_comparison_halfcheetah.sh

Results

The result will be available in the log dir specified.

The correlation between the predicted reward and the ground-truth reward tested on the unseen_trajs is reported at the end of running on console, or, if you are using the bash script, at the end of the d_rex.log or ssrr.log.

We also provide our run results available in results/{halfcheetah/hopper/ant}/{airl/noisy_airl}_data_ssrr_{1/2/3/4/5}/{drex/ssrr}/.

Code Structure

SSRR is implemented in script_experiment/ssrr.py, Agents/SSRRAgent.py, Datasets/NoiseDataset.py.

D-REX is implemented in script_experiment/drex.py, scrip_experiment/drex_utils.py, and script_experiment/tf_commons/ops.

Both implementations are adapted from https://github.com/dsbrown1331/CoRL2019-DREX.

Policy Learning

We utilize stable-baselines to optimize policy over the reward we learned.

Running

Before running, you should edit script_experiment/rl_utils/sac.yml to change the learned reward model directory, for example:

  env_wrapper: {"script_experiment.rl_utils.wrappers.CustomNormalizedReward": {"model_dir": "/home/zac/Programming/Zac-SSRR/results/halfcheetah/noisy_airl_data_ssrr_4/ssrr/", "ctrl_coeff": 0.1, "alive_bonus": 0.0}}

Examples of running command would be

python script_experiment/train_rl_with_learned_reward.py \
 --algo=sac \
 --env=HalfCheetah-v3 \
 --tensorboard-log=./results/HalfCheetah_custom_reward/ \
 --log-folder=./results/HalfCheetah_custom_reward/ \
 --save-freq=10000

Please note the flag --env-kwargs=terminate_when_unhealthy:False is necessary for Hopper and Ant as discussed in our paper Supplementary D.1.

Examples of running evaluation the learned policy's ground-truth reward would be

python script_experiment/test_rl_with_ground_truth_reward.py \
 --algo=sac \
 --env=HalfCheetah-v3 \
 -f=./results/HalfCheetah_custom_reward/ \
 --exp-id=1 \
 -e=5 \
 --no-render \
 --env-kwargs=terminate_when_unhealthy:False

Results

The result will be available in the log folder specified.

We also provide our run results in results/.

Code Structure

The code script_experiment/train_rl_with_learned_reward.py and utils/ call stable-baselines library to learn a policy with the learned reward function. Note that utils could not be renamed because of the rl-baselines-zoo constraint.

The codes are adjusted from https://github.com/araffin/rl-baselines-zoo.

Random Seeds

Because of the inherent stochasticity of GPU reduction operations such as mean and sum (https://github.com/tensorflow/tensorflow/issues/3103), even if we set the random seed, we cannot reproduce the exact result every time. Therefore, we encourage you to run multiple times to reduce the random effect.

If you have a nice way to get the same result each time, please let us know!

Ending Thoughts

We welcome discussions or extensions of our paper and code in Issues!

Feel free to leave a star if you like this repo!

For more exciting work our lab (CORE Robotics Lab in Georgia Institute of Technology led by Professor Matthew Gombolay), check out our website!