Denoised MDPs: Learning World Models Better Than The World Itself

Tongzhou Wang, Simon S. Du, Antonio Torralba, Phillip Isola, Amy Zhang, Yuandong Tian

We provide a PyTorch implementation of Denoised MDPs: Learning World Models Better Than The World Itself, published in ICML 2022.

• arXiv
• Project Page

(We also provide a PyTorch implementation of Dreamer that is carefully written and verified to reproduce results. See here for usages.)

The raw real world is noisy. How can reinforcement learning agent successfully learn with such raw data, where signals can be strongly entangled with noises? Denoised MDP characterizes information into four distinct types, based on controllability and relation with rewards, and proposes to extract a state representation space containing only information both controllable and reward-relevant. Under this view, several prior works can be seen as insufficiently removing noisy information.

To properly extract only the useful signal, Denoised MDP considers novel factorized MDP transition structures, where signal representation and noise representation are separated into distinct latent spaces. The state abstraction (i.e., representation learning) problem is turned into a regularized model fitting problem: fitting the factorized forward model to collected trajectories, while requiring the signal latents to be minimally informative of the raw observations.

The resulting variational formulation (derivation in paper) successfully disentangles a variety of noise types (and also noiseless settings), outperforming baseline methods that often can only do well for certain particular noise types.

Visualizations

For environments with distinct types of noises, we visualize latent factorization idenfitied by Denoised MDP, and other baseline methods. Only Denoised MDP successfully disentangle signal from noises across all environments.

• Task: Press green button to shift TV hue to green (a RoboDesk variant).
  True Signal: Robot joint position, TV green-ness, Green light on desk.
  True Noise: Lighting, Camera, TV content, Imperfect sensor.

  https://user-images.githubusercontent.com/5674597/173155667-d4bcc7af-1f12-4ba3-a733-ef9d5f631c96.mp4

  • Task: Make reacher robot touch the target red object.
    True Signal: Robot joint position, Target location.
    True Noise: Background.

    https://user-images.githubusercontent.com/5674597/173155695-d85f5866-0bbe-4451-9898-d032d3bcf51c.mp4

  • Task: Make the walker robot move forward when sensor readings are noisily affected by background images.
    True Signal: Robot joint position.
    True Noise: Background, Imperfect sensor.

    https://user-images.githubusercontent.com/5674597/173155705-02f098f7-dca8-4022-995a-57b1aa854935.mp4

  • Task: Move half cheetah robot forward when camera is shaky.
    True Signal: Robot joint position.
    True Noise: Background, Camera.

    https://user-images.githubusercontent.com/5674597/173155710-285d0088-b3a3-42e7-9127-a177e4f6b955.mp4

  Requirements

  The code has been tested on

  • CUDA 11 with NVIDIA RTX Titan, NVIDIA 2080Ti, and NVIDIA Titan XP,
  • mujoco=2.2.0 with egl renderer.

  Software dependencices (also in requirements.txt):

  torch>=1.9.0
  tqdm
  numpy>=1.17.0
  PIL
  tensorboardX>=2.5
  attrs>=21.4.0
  hydra-core==1.2.0
  omegaconf==2.2.1
  mujoco
  dm_control

  Environments

  The code supports the following environments:

  kind	spec	Description
  robodesk	${TASK_NAME} or ${TASK_NAME}_noisy (e.g., tv_green_hue_noisy)	RoboDesk environment (96x96 resolution) with a diverse set of distractors (when using ${TASK_NAME}_noisy variant). The distractors are implemented and descriped in details at this RoboDesk fork.
  dmc	${DOMAIN_NAME}_${TASK_NAME}_${VARIANT} with VARIANT being one of [noiseless, video_background, video_background_noisy_sensor, video_background_camera_jitter] (e.g., cheetah_run_video_background_camera_jitter)	DeepMind Control (DMC) environment (64x64 resolution) with four possible variants, representing different types of noises.

  In the paper, we used the following 13 environments, all run with 1000 max episode length and action repeat of 2:

  kind	spec
  robodesk	tv_green_hue_noisy
  dmc	${DOMAIN_NAME}_${TASK_NAME} in [cheetah_run, walker_walk, reacher_easy] each with all 4 VARIANT options

  NOTE: All noisy environments require the driving_car class of the Kinetics-400 training dataset. Some instructions for downloading the dataset can be found here. After downloading, you may either place it under ~kinetics/070618/400 (so that the videos are ~kinetics/070618/400/train/driving_car/*.mp4) or specify KINETICS_DIR environment variatble (so that the videos are ${KINETICS_DIR}/train/driving_car/*.mp4).

  Training and Evaluation

  env CUDA_VISIBLE_DEVICES=0 \              # GPU ID for training
      EGL_DEVICE_ID=0 \                     # GPU ID for rendering
      KINETICS_DIR=/path/to/kinectics/ \    # Videos for noisy env
      python main.py \
          env.kind=robodesk \               # Env kind
          env.spec=tv_green_hue_noisy \     # Env spec
          learning.model_learning.kl.alpha=2 \       # alpha, weight of the KL terms
          learning.model_learning.kl.beta_y=0.125 \  # beta, smaller => stronger regularization
          learning.model_learning.kl.beta_z=0.125 \  # beta, smaller => stronger regularization
          seed=12 \                         # Seed
          output_folder=subdir/for/output/  # [Optional] subdirectory under `./results`
                                            # for storing outputs. If not given, a folder name will
                                            # be automatically constructed with information from
                                            # given config

  Hyperparameter choice (see also Appendix A.2 for more details):

  • alpha parameter is selected proportional to the size of observation. For DMC (64x64x3 observations), we use alpha=1. For RoboDesk (96x96x3 observations), we use alpha=2.
  • beta parameter (for y and z component) controls the regularization strength, and should be set in (0, 1). Noisier environments benefit from a smaller value.

  Default behaviors:

  • Train the Figure 2b Denoised MDP variant over 10^6 environment steps, with 5000 steps prefilling the replay buffer, and then training for 100 iterations for every 100 steps.
  • Optimize policy bybackpropagation through dynamics (Dreamer-style). One can switch to Soft Actor-Critic via specifying learning/policy_learning=sac (note the / rather than .).
  • Evaluate for 10 episodes every 10000 steps.
  • Visualize for 3 episodes (both full reconstruction and with noise latent fixed) every 20000 steps.

  We use Hydra to handle argument specification. You can use Hydra's overriding syntax to specify all sorts of config options. See config.py for the complete set of options. Additionally, one can also check the config.yaml file generated in output directory for all options.

  Figure 2c Variant with x, y and z Latents

  To use the Figure 2c variant with three sets of latents x, y, and z, set

  learning.model.transition.z.belief_size={NONZERO_Z_BELIEF_SIZE} \
  learning.model.transition.z.state_size={NONZERO_Z_STATE_SIZE} \

  for some non-zero belief and state sizes for the z latent component.

  In the paper Appendix B.5, a variant similar to Figure 2c is compared, with z prior not depending on y. To reproduce those results on DeepMind Control Suite environments, set

  learning.model.transition.x.belief_size=120 \
  learning.model.transition.x.state_size=20 \
  learning.model.transition.y.belief_size=70 \
  learning.model.transition.y.state_size=10 \
  learning.model.transition.z.belief_size=70 \
  learning.model.transition.z.state_size=10 \
  learning.model.transition.z_prior_uses_y=False \

  Reproducing Dreamer

  When y and z latent spaces are completely turned off (i.e., empty), the code essentially is Dreamer. This can be done by setting

  learning.model.transition.x.belief_size=200 \  # give `x` the dimensionality specified in Dreamer paper
  learning.model.transition.x.state_size=30 \
  learning.model.transition.y.belief_size=0 \
  learning.model.transition.y.state_size=0 \
  learning.model.transition.z.belief_size=0 \
  learning.model.transition.z.state_size=0 \

  Code Structure

  To facilitate easier parsing and usage of this repository, we provide a detailed note on how our code is structured here.

  Pre-emption

  Upon receiving SIGUSR1, the provided code starts writing all necessary states (including replay buffer) into a folder under the output directory (usually taking up to 10 minutes), and exits naturally afterwards. When the code is run with the same output directory, it continues from that state (and deletes the saved state). This may be particularly useful if you are running on a shared cluster.

  Citation

  Tongzhou Wang, Simon S. Du, Antonio Torralba, Phillip Isola, Amy Zhang, Yuandong Tian. "Denoised MDPs: Learning World Models Better Than The World Itself" International Conference on Machine Learning. 2022.

  @inproceedings{wang2022denoisedmdps,
    title={Denoised MDPs: Learning World Models Better Than The World Itself},
    author={Wang, Tongzhou and Du, Simon S. and Torralba, Antonio and Isola, Phillip and Zhang, Amy and Tian, Yuandong},
    booktitle={International Conference on Machine Learning},
    organization={PMLR},
    year={2022}
  }

  If you find the RoboDesk distractor options (see this repository for more options and details) useful for your research, please also cite the following:

  Click to show RoboDesk distractor bibtex!

  ``` @misc{wang2022robodeskdistractor, author = {Tongzhou Wang}, title = {RoboDesk with A Diverse Set of Distractors}, year = {2022}, howpublished = {\url{https://github.com/SsnL/robodesk}}, } @misc{kannan2021robodesk, author = {Harini Kannan and Danijar Hafner and Chelsea Finn and Dumitru Erhan}, title = {RoboDesk: A Multi-Task Reinforcement Learning Benchmark}, year = {2021}, howpublished = {\url{https://github.com/google-research/robodesk}}, } ```

  Questions

  For questions about the code provided in this repository, please open an GitHub issue.

  For questions about the paper, please contact Tongzhou Wang (tongzhou _AT_ mit _DOT_ edu).

  License

  This repo is under CC BY-NC 4.0. Please check LICENSE file.