LocoMuJoCo v1.0.1

🚀 New LocoMuJoCo Version 🚀

We are thrilled to announce the latest release of our package, packed with enhanced features and key advantages that elevate your work with humanoid and quadruped environments to new heights. Here's what you can look forward to:

New Features:

• MuJoCo and MJX Support: Enjoy seamless integration with MuJoCo for single environments and MJX for parallel environments, offering unmatched flexibility.
• Efficient JAX Algorithms: Utilize clean, single-file JAX algorithms for PPO, GAIL, AMP, and DeepMimic, enabling quick and efficient benchmarking.
• Expanded Environment Library: Explore a new collection of 12 humanoid and 4 quadruped environments.
• Lightning-Fast Training: Experience unparalleled training speeds with our combined training and environment JIT-compiled function.
• Extensive Motion Capture Datasets: Access over 22,000 motion capture datasets from AMASS, LAFAN1, and native LocoMuJoCo, all retargeted for each humanoid environment.
• Robot-to-Robot Retargeting: Easily retarget any existing dataset from one robot to another. Add your humanoid and get immediate access to all +22,000 datasets!
• Advanced Trajectory Metrics: Leverage powerful trajectory comparison metrics, including dynamic time warping and discrete Fréchet distance, all implemented in JAX.
• Built-In Randomization: Enhance your simulations with built-in domain and terrain randomization, ensuring robust and adaptable models.
• Modular Design: Customize your workflow with a modular design that allows you to define, swap, and reuse components like observation types, reward functions, terminal state handlers, and domain randomization.

LocoMuJoCo v0.4.1

Added MyoLab's MyoSkeleton Model

ssstwitter.com_1726241460284.mp4

• 151 DoF Humanoid model with realistic spine and neck.
• physiologically realistic model.
• amongst the most comprehensive models to date.

Remarks on model

• this environment is still under active development.
• we are still working on a working policy, so there is no imitation learning script setup yet. But you can modify the existing experiment files by taking the hyperparameters of the other environments to try it out!

Bug Fixes:

• #28
• #30
• #39
• #40
• #41
• #45

Credits

Many thanks to the MyoLab team for providing and helping integrate this model into LocoMuJoCo!

LocoMuJoCo v0.3.0

Added Unitree G1 Humanoid.

• included environment, documentation, and imitation learning scripts.
• fixed bug in testing
  • mujoco is fixed now
  • updated test datasets

LocoMuJoCo v0.2.2

Small fix in datasets.

LocoMuJoCo v0.2.0

Changelog

• New Features:
  • Added full documentation.
  • Added dataset generation pipeline.
  • Added imitation learning experiment script with hyperparameters.
• Datasets:
  • added perfect datasets for Atlas.carry and HumandTorque4Ages tasks.
  • fixed issue with running datasets.
• Fixes:
  • Fixed custom reward interfaces for robot environments (#15).
  • Fixed perfect dataset loader for UnitreeA1 environment (#20).
  • Added render modes to Gymansium interface (#14).

LocoMuJoCo v0.1.0

Initial release of LocoMuJoCo.

This includes:

• 12 environments
• 27 tasks, each represented by different datasets
• for the majority of the tasks, perfect datasets (ground-truth states and actions) are also available
• easy dynamics randomization
• Gym and Mushroom-RL interfaces
• interface for custom reward
• examples and baselines