AngoraPy 0.10.8

Anthropomorphic Task API

• Integrated new, adapted version of the MuJoCo Menagerie ShadowHand. Old policies trained on the previous hand model might not work with this newer model.
• Added dm_control dependency to programmatically build MuJoCo models. This does not replace the gym-basis of the API but extends it. However, while gym(nasium) will in the foreseeable future remain to be the standard by which Environments are communicated with in AngoraPy, we will move many interactions with MuJoCo to dm_control in the future. This should, however, not affect normal usage of AngoraPy.
• AnthropomorphicEnv can now accept a wider range of definitions of models of the agents body.
• Refactored methods of AnthropomorphicEnv and BaseShadowHandEnv such that methods that can be made use of in any anthropomorphic/hand task moved up the inheritance hierarchy from more specific classes.
• Added getters for senses to AnthropomorphicEnv and a default _get_obs()

Modeling API

• make_input_layers() now returns a dictionary mapping the name of the modality to the input layer
• Model registration now is achieved via decorators (@register_model('model_name'))

Backend

• Upgraded to TensorFlow 2.15.1, thus cuda will now be installed with AngoraPy (via TF)
• Improved model optimization during training.
• Fixed state resets during optimization and inference.

Deployment

• Updated Dockerfiles.

Misc

• Improved (experimental) pretraining.

Documentation

• We now maintain a list of major features in the README, where we will also include future features that we currently work on.