PyBullet

PyBullet
Name	PyBullet
Developer	Carnegie Mellon University Robotics Institute; later Google
Initial release	2015
Programming language	Python (programming language), C++
Operating system	Linux, Microsoft Windows, macOS
Genre	Physics engine, Robotics simulator

PyBullet PyBullet is an open-source Python module for physics simulation and robotics research. It integrates a rigid-body dynamics engine with collision detection, kinematics, and visualization, and has been used in research at institutions such as Massachusetts Institute of Technology, Stanford University, Harvard University, University of California, Berkeley, and University of Oxford. Developers and researchers from Google and the Carnegie Mellon University Robotics Institute have contributed to its codebase and ecosystem.

Overview

PyBullet provides bindings to a real-time physics engine originally developed in C++ and exposed to Python (programming language), facilitating rapid prototyping for teams at OpenAI, DeepMind, Facebook AI Research, Microsoft Research, and laboratories at NASA and Toyota Research Institute. The project interoperates with robotics platforms and middleware like ROS and supports model formats used by Universal Robots, ABB, KUKA, and research in humanoid platforms such as Atlas (robot), ASIMO, NAO (robot), and PR2. PyBullet is used alongside machine learning frameworks such as TensorFlow, PyTorch, JAX, and Scikit-learn for reinforcement learning and imitation learning experiments.

Features

PyBullet implements collision detection, convex decomposition, continuous collision detection, constraint solvers, and soft body dynamics useful to projects at MIT Media Lab, ETH Zurich, Imperial College London, and industry labs like NVIDIA and Intel. It supports visualization via OpenGL and provides client-server APIs that integrate with tools such as Gazebo, Unity (game engine), Unreal Engine, and development environments like Visual Studio Code. Asset importers handle formats such as URDF, SDF, and MJCF used by Boston Dynamics research, Honda robotics studies, Yale University tactile sensing labs, and the Allen Institute for AI. PyBullet exposes inverse kinematics, forward dynamics, joint control, and trajectory optimization routines applied in projects at Caltech, University of Toronto, and ETH Zurich.

Architecture and Implementation

Under the hood, the engine uses optimized data structures and algorithms from computational geometry and numerical linear algebra developed in C++ and exposed through SWIG or custom bindings for Python (programming language). The core leverages broadphase collision detection, narrowphase algorithms, and iterative solvers inspired by academic work from Stanford University and University of Pennsylvania. It integrates with visualization backends that rely on OpenGL and windowing systems such as X Window System and Wayland on Linux, as well as Microsoft Windows APIs and macOS frameworks. The codebase has evolved through contributions from researchers affiliated with Cornell University, University of Michigan, Delft University of Technology, and corporate research groups at Amazon, Sony CSL, and Siemens.

Use Cases and Applications

Researchers employ PyBullet for reinforcement learning benchmarks influenced by tasks from Atari 2600 suites, continuous control benchmarks from OpenAI Gym, and imitation learning datasets produced in collaboration with Carnegie Mellon University and Stanford University. Industrial applications include robot arm manipulation studies for companies like ABB and KUKA, autonomous vehicle sensor simulation for Waymo and Uber ATG-style projects, and biomechanical analysis in labs such as Johns Hopkins University and UCSF. Educational courses at MIT, Stanford University, and University of Cambridge use PyBullet to teach robotics lab exercises, while workshops at conferences like ICRA, RSS, NeurIPS, ICLR, and CVPR feature PyBullet-based demonstrations.

Comparison with Other Simulators

Compared to Gazebo, PyBullet emphasizes lightweight Python integration and rapid iteration similar to projects at OpenAI and DeepMind, whereas MuJoCo historically provided high-performance optimization and was used in research at DeepMind and Stanford University before changes in its licensing. Versus DART (software), PyBullet offers simpler setup for Python users; compared to Chrono::Engine, it prioritizes robotics convenience over multibody vehicle dynamics used by Argonne National Laboratory. In visualization and game-engine fidelity, PyBullet bridges workflows used with Unity (game engine) and Unreal Engine while remaining more accessible for researchers from EPFL and KAIST.

Development and Community

The project’s community includes contributors from academia and industry, pull requests and issue discussions involving developers from Google, Carnegie Mellon University, OpenAI, NVIDIA, and universities such as MIT, Stanford University, ETH Zurich, and University of Cambridge. Community resources and tutorials are shared in forums frequented by members of Stack Overflow, mailing lists associated with ROS, and workshops at conferences like ICRA and NeurIPS. Third-party integrations and forks have been developed by teams at Amazon, Microsoft Research, Facebook AI Research, and academic labs at University of Toronto and University of Oxford.

Licensing and Availability

PyBullet is distributed under an open-source license and is available through package managers used by developers on Linux, Microsoft Windows, and macOS. Commercial users and research labs at Google and Toyota Research Institute have adopted it for prototyping, while academic groups at Harvard University, Yale University, and Princeton University rely on it for reproducible experiments. Binary wheels and source code are maintained for compatibility with Python (programming language) versions and integrations with machine learning libraries such as TensorFlow and PyTorch.

Category:Physics engines