MooJoCo provides a unified framework for implementing and interfacing with MuJoCo and MuJoCo-XLA simulation environments. The main goal of this framework is to unify the development and interfaces of environments implemented in native MuJoCo (MJC) and MuJoCo-XLA (MJX).

MJCFMorphology and MJCFArena

• Both are parameterized/reconfigurable MJCF (MuJoCo-XML) generators that serve as input to an environment.
• MJCFMorphology defines the robot morphology.
  • MJCFMorphologies are parameterized using the Framework for Parameterized Robot Specifications (FPRS).
  • MJCFMorphologies follow a modular design, dividing the robot into distinct parts.
• MJCFArena defines the arena in which the robot is places (i.e. all non-morphological structures).
  • MJCFArena are reconfigurable via a ArenaConfiguration
    • reconfigurable via a configuration requirements.txt

Unified MJC and MJX environment interface

• Reconfigurable through an environment configuration
• Functional programming
  • MJX requires the functional programming paradigm. To unify MJC and MJX, we thus apply this stronger coding constraint to the MJC side as well.
  • Simply said: functions are pure, and should not have side effects.
  • Environments expect states and return states.
• Gymnasium-like interface
  • Environments provide an observation and action space (automatically)
  • Outside code interacts with the environment using the typical step, reset, render and close functions.
  • Main difference with the original gymnasium interface: here the environment returns and expects states, which encapsulates the original observations, rewards, info, etc. in a datastructure, together with the underlying physics state.
• Differences between MJCEnv (i.e. a native MuJoCo environment) and MJXEnv (i.e. an MJX environment):
  • development: MJXEnv's should be implemented in JAX, MJCEnv's do not necessarily require JAX.
  • usage: MJXEnvState contains JAX arrays (and the additional mjx.Model and mjx.Data structures), the MJCEnvState uses numpy arrays.
• Observations are implemented using Observables; these define retriever functions that extract the observation based on the state datastructure and provide an easy way to implement observations.
• DualMuJocoEnvironment provides the unification of MJC and MJX, and allows conditional environment creation based on a backend string.
• Both MJC and MJX support human-mode and rgb_array rendering modes.
  • Note: MJX rendering is slow due to the offloading of datastructures of the GPU

Examples

For practical applications and demonstrations of MooJoCo, please refer to the Bio-inspired Robotics Testbed, which employs this framework extensively.

Installation

pip install moojoco