Skip to content
This repository has been archived by the owner on Sep 1, 2024. It is now read-only.

Latest commit

 

History

History
94 lines (84 loc) · 5.29 KB

CHANGELOG.md

File metadata and controls

94 lines (84 loc) · 5.29 KB
Raw

Changelog

main (v0.2.0)

Breaking changes

  • Migrated from gym to Gymnasium
  • gym==0.26.3 is still required for the dm_control and pybullet-gym environments
  • Transition and TransitionBatch now support the terminated and truncated booleans instead of the single done boolean previously used by gym
  • Migrated calls to env.reset() which now returns a tuple of obs, info instead of just obs
  • Migrated calls to env.step() which now returns a observation, reward, terminated, truncated, info
  • Migrated to Gymnasium render API, environments are instantiated with render_mode=None by default
  • DMC and PyBullet envs use the original gym wrappers to turn them into gym environments, then are wrapper by gymnasium.envs.GymV20Environment
  • All Mujoco envs use the DeepMind Mujoco bindings, mujoco-py is deprecated as a dependency
  • Custom Mujoco envs e.g. AntTruncatedObsEnv inherit from gymnasium.envs.mujoco_env.MujocoEnv, and access data through self.data instead of self.sim.data
  • Mujoco environment versions have been updated to v4 fromv2 e.g. Hopper-v4
  • Fixed PlaNet to save model to a directory instead of a file name
  • Added follow-imports=skip to mypy CI test to allow for gymnasium/gym wrapper compatibility
  • Bumped black to version 0.23.1 in CI

v0.2.0.dev4

Main new features

Breaking changes

  • Model.reset() and Model.sample() signature has changed. They no longer receive TransitionBatch objects, and they both return a dictionary of strings to tensors representing a model state that should be passed to sample() to simulate transitions. This dictionary can contain things like previous actions, predicted observation, latent states, beliefs, and any other such quantity that the model need to maintain to simulate trajectories when using ModelEnv.
  • Ensemble class and sub-classes are assumed to operate on 1-D models.
  • Checkpointing format used by save() and load() in classes GaussianMLP and OneDTransitionRewardModel changed, making old checkpoints incompatible with the new version.
  • use_silu argument to GaussianMLP has been replaced by activation_fn_cfg, which is an omegaconf.DictConfig specifying the class to use for the activation functions, thus giving more flexibility.
  • Removed unnecessary nesting inside dynamics_model Hydra configuration.
  • SAC agents prior to v0.2.0 cannot be loaded anymore.

Other changes

  • Added add_batch() method to mbrl.util.ReplayBuffer.
  • Added functions to mbrl.util.models to easily create convolutional encoder/decoders with a desired configuration.
  • mbrl.util.common.rollout_agent_trajectories now allows rolling out a pixel-based environment using a policy trained on its corresponding non-pixel environment version.
  • ModelTrainer can be given eps for Adam optimizer. It now also includes a progress bar using tqdm (can be turned off).
  • CEM optimizer can now be toggled between using clipped normal distribution or truncated normal distribution.
  • mbrl.util.mujoco.make_env can now create an environment specified via an omegaconf configuration and hydra.utils.instantiate, which takes precedence over the old mechanism if both are present.
  • Fixed bug that assigned wrong termination functino to humanoid_truncated_obs env.

v0.1.4

  • Added MPPI optimizer.
  • Added iCEM optimizer.
  • control_env.py now works with CEM, iCEM and MPPI.
  • Changed algorithm configuration so that action optimizer is passed as another config file.
  • Added option to quantize pixel obs of gym mujoco and dm control env wrappers.
  • Added a sequence iterator, SequenceTransitionSampler, that always returns a fixed number of random batches.

v0.1.3

  • Methods loss, eval_score and update of Model class now return a tuple of loss/score and metadata. Currently, supports the old version as well, but this will be deprecated in v0.2.0.
  • ModelTrainer now accepts a callback that will be called after every batch both during training and evaluation.
  • Normalizer in util.math can now operate using double precision. Utilities now allow specifying if replay buffer and normalizer should use double or float via Hydra config.

v0.1.2

  • Multiple bug fixes
  • Added a training browser to compare results of multiple runs
  • Deprecated ReplayBuffer.get_iterators() and replaced with mbrl.util.common.get_basic_iterators()
  • Added an iterator that returns batches of sequences of transitions of a given length

v0.1.1

  • Multiple bug fixes
  • Added third_party folder for pytorch_sac and dmc2gym
  • Library now available in pypi
  • Moved example configurations to package mbrl.examples, which can now be run as python -m mbrl.examples.main, after pip installation

v0.1.0

Initial release