RaiSim

A multi-body physics engine for robotics and reinforcement learning, known for fast and accurate contact solving.

RaiSim was developed by Jemin Hwangbo and colleagues, and is now maintained by RaiSim Tech. Its core comes from a 2018 IEEE RA-L paper by Hwangbo, Lee, and Hutter describing a “per-contact iteration” method that iterates over contact points one at a time and uses bisection to solve for contact force, which the paper reports as roughly twice as fast as two existing methods. ETH Zurich's 2019 paper on training the ANYmal quadruped with reinforcement learning used this same contact solver in its custom simulator. RaiSim focuses on simulating rigid and articulated bodies fast and accurately on the CPU, offering a C++ interface, Python bindings called raisimPy, and reinforcement-learning examples in raisimGymTorch; later versions added soft bodies, granular media, tendons, and camera/depth sensor rendering. The older raisimLib repository was archived in April 2026 and replaced by the binary-distributed RaiSim2, which reached v2.7.0 in September 2026. It is not open source: an academic license is free, while a commercial license costs $1,500 per person per year.

ExampleUsing raisimGymTorch, run many quadruped-robot environments in parallel on one computer to train a reinforcement-learning policy for ANYmal to walk over height-map terrain.

Also called

RaiSim2, raisimLib

Related

Physics Engine · MuJoCo (Multi-Joint dynamics with Contact) · Isaac Gym · RL-based Locomotion Control · ANYbotics ANYmal · ETH Zurich Robotic Systems Lab

Sources

RaiSim 文档：License（v2.7.0） (Chinese)
Per-Contact Iteration Method for Solving Contact Dynamics (IEEE RA-L 2018)
raisimTech/raisim2Lib GitHub

As of

2026-09