Jiminy is a fast, physically accurate, and portable simulator for poly-articulated robotic systems, designed for reinforcement learning research. It offers a C++ core with Python bindings, enabling efficient simulation of complex robot dynamics and seamless integration with OpenAI Gym for machine learning tasks.

How It Works

Jiminy leverages Pinocchio for minimal coordinate and Lagrangian dynamics, providing near-analytical solutions for articulated systems. This approach prioritizes speed and numerical stability. It includes a contact model and constraint solver based on energy dissipation, and supports flexible controller and sensor timing, making it suitable for demanding ML simulations.

Quick Start & Requirements

• Install Jiminy: python -m pip install --prefer-binary jiminy_py[meshcat,plot]
• Install Gym Jiminy: python -m pip install --prefer-binary gym_jiminy[all]
• Requirements: Python 3.10+, Linux, macOS, Windows. Pre-compiled binaries are available on PyPI.
• Documentation: https://duburcqa.github.io/jiminy/

Highlighted Details

• GPU-accelerated offscreen rendering via Panda3d.
• Interactive 3D viewer with Meshcat for remote environments (Colab, Jupyter).
• Modular learning environment pipeline with plug-and-play components.
• Simulation logs exportable to binary or HDF5.
• Supports simulating multiple interacting articulated systems.

Maintenance & Community

• Developed by the Robotics Learning and Interaction (RLI) team at INRIA.
• Active development and community support channels are not explicitly mentioned in the README.

Licensing & Compatibility

• License: Apache 2.0.
• Compatible with most RL frameworks (Stable Baselines 3, Tianshou, Rllib) implementing the standard OpenAI Gym API.

Limitations & Caveats

The README mentions "primitives and meshes yet to come" for contact collision, suggesting mesh-based collision is still under development. Detailed performance benchmarks are not provided.