DART: A Research-Oriented Physics Engine for Robotics and Animation

DART is an open-source C++20 physics engine designed specifically for robotics, animation, and machine learning research. It uses generalized coordinates and the Featherstone algorithm to provide transparent kinematics and dynamics calculations, supporting multiple model formats such as URDF, SDF, and MJCF.

• Original Authors/Maintainers: The dartsim organization (key contributors include Jeongseok Lee, Michael X. Grey, Sehoon Ha, etc.)
• Source Platform: GitHub
• Original Title: DART: Dynamic Animation and Robotics Toolkit
• Original Link: https://github.com/dartsim/dart
• Release Status: Continuously updated; the current active branch is DART 7

DART (Dynamic Animation and Robotics Toolkit) is a research-focused physics engine written in C++20 with Python bindings. Unlike many "black-box" simulators, DART provides researchers with transparent access to kinematics, dynamics, collision detection, and constraint solving.

The project's core design philosophy uses generalized coordinates to describe articulated rigid body systems and Featherstone's Articulated Body Algorithm (ABA) for accurate and stable motion dynamics calculations. This approach is widely regarded as the optimal balance between computational efficiency and numerical stability in robotics and computer animation.

DART's core advantage lies in its research-oriented design. It is not just a simulation tool but a dynamics calculation library:

• Featherstone Algorithm: Uses the classic articulated body algorithm to ensure efficiency and accuracy in multi-body system dynamics calculations
• Generalized Coordinate System: Uses minimal coordinate representation to avoid redundant computations and constraint drift issues
• Direct Access to Dynamic Quantities: Researchers can directly obtain intermediate calculation results such as mass matrices, Coriolis forces, and gravity terms

DART provides a unified model loading API supporting multiple industry-standard formats:

• URDF: The most commonly used robot description format in the ROS ecosystem
• SDF: Native format for the Gazebo simulator
• MJCF: XML configuration format for MuJoCo
• SKEL: DART's own skeleton format

This multi-format support allows researchers to easily migrate existing projects without re-modeling.

DART's design considers future computational architecture evolution:

• Current: Cross-platform CPU support, running on Linux, macOS, and Windows
• Roadmap: Multi-core parallelism, SIMD instruction set optimization, accelerator backends (GPU/TPU)

DART is widely used in the research and validation of robot control algorithms:

• Motion Planning: Trajectory optimization using accurate dynamic models
• Reinforcement Learning: Serving as a physical simulation backend for training robot policies
• Human-Robot Interaction: Simulating safe interactions between robots and human environments

In the animation field, DART provides physically accurate character animation generation capabilities:

• Character Controller Development: Physics-based animation systems
• Cloth and Soft Body Simulation: Extending rigid body systems to soft structures
• Motion Retargeting: Adapting captured data to different characters