Dense foliage
Wind-blown trees and grass that cast real shadows, so outdoor perception sees clutter it will meet in the field.
Skip to content
log scale
v2.10.3Read the release notes
Train robotsfaster than reality.
Quadrupedal locomotion, learned in 10 seconds. Contact-rich robot physics hundreds of times faster than real time, with photoreal cameras and LiDAR.
10 s
to learn quadrupedal locomotion
A walking quadruped trained from scratch with raisimGymTorch and PPO
334×
faster than real time
A full day of ANYmal walking on rough terrain, simulated in 4.3 minutes
25.7×
faster LiDAR than the competition
64-beam spinning LiDAR over a 200 m terrain, for perception in the loop
Free
for academic research
Academic licenses cost nothing, and a one-month trial is free for everyone
Robot learning
From simulation to a trained policy.
A quadruped learns to walk in 10 seconds. Build environments in C++ or Python, train with PPO in PyTorch, render camera observations headless, and watch any run live, all on the same RaiSim world.
Observations
rayrai
Render camera and LiDAR observations headless for training, or embed photoreal views in your own tools.
Debugging
TCP viewer
Watch a running simulation live, inspect contacts, and pause, step or drag the robot.
Reinforcement learning
raisimGymTorch
Vectorized environments and PPO in PyTorch: quadrupedal locomotion in 10 seconds, tens of billions of steps on one desktop.
Python
RaisimPy
The full C++ API from Python with NumPy arrays, for prototyping scenes and controllers.
Performance
More experience per hour. Faster policies.
Reinforcement learning needs billions of steps. RaiSim's lean solvers give you more of them from the hardware you already have: a quadruped learns to walk in 10 seconds, so you iterate on rewards and controllers while the idea is still fresh.
64 quadrupedsStepping on rough terrain, all in one world
Speedup over the competition
The same robotics workloads on the same machine; higher is better.
Chain, 10 links3.78×
Chain, 20 links6.96×
ANYmal standing2.56×
ANYmal falling2.61×
ANYmal on heightmap4.81×
Primitive contacts2.36×
Strandbeest2.74×
Heightmap LiDAR25.7×
Ray queries147×
Deformable cloth11.9×
AMD Ryzen 9 3950X, median of three runs, October 2026. The deformable cloth row compares different deformable models. Workloads and method
Perception
Train on images that look real.
rayrai, RaiSim's renderer, turns the running simulation into RGB, depth and LiDAR in the same process, headless on a training server or in a window, with the lighting, weather and materials your robot will meet outside.
Explore rayraiGlass and refraction
Glass and transparent objects refract light the way a camera sees them, a classic failure case for depth and grasping.
Weather
Rain, fog, snow and time of day, for policies that hold up in bad weather.
Photoreal scenes
Cities, forests and warehouses with physically based materials and lighting.
LiDAR point clouds
Spinning LiDAR that scans the same world the robot walks in, ready for mapping and navigation.
Depth camera
Depth aligned with the RGB camera, for terrain perception and grasping.
RGB camera
Robot-mounted cameras rendered headless on a server, with sensor noise.
Physics
Physics that holds up on real robots.
Legged-robot controllers trained in RaiSim have walked real robots over rough terrain in published research. Hard contacts with friction, closed loops, tracks, deformable soil and granular media are solved together, and the same inputs give the same trajectory, run after run.
Articulated vehicles
Suspension, steering and drive shafts move as one articulated system, with live spring and damper forces.
Deformable terrain
Tires sink and leave lasting ruts, with the whole contact patch resolved, for off-road and planetary rovers.
Closed kinematic loops
Linkage-driven legs and parallel mechanisms: a Strandbeest walks on 36 closed loops from a single crank.
Dexterous contact
A Shadow Hand from OpenUSD grasps a cube, with many simultaneous contacts for in-hand manipulation.
Bring your robots
URDF, MJCF and OpenUSD models load directly, with their meshes and materials.
Tracked vehicles
Tracks are simulated as driven belts, so a flipper robot crosses loose rubble and climbs stairs on contact friction alone.
Hall of fame
Research that ran on RaiSim.
Controllers trained in RaiSim have run a marathon, raced across beach sand and hiked alpine trails. These are the papers in Nature and Science Robotics, and more, whose robots learned in RaiSim.
Nature2026
A quadruped robot designed to complete a marathon on a single battery charge
RAIBO2 ran a full marathon on one battery charge, in 4 hours 19 minutes.
Lee et al. · KAISTScience Robotics · 2025
High-speed control and navigation for quadrupedal robots on complex and discrete terrain
Raibo runs on vertical walls, jumps a 1.3 m gap and crosses stepping stones at 4 m/s.
Kim et al. · KAISTScience Robotics · 2023Learning quadrupedal locomotion on deformable terrain
Raibo runs on soft beach sand at 3.03 m/s, its feet buried at every step.
Choi et al. · KAISTScience Robotics · 2022Learning robust perceptive locomotion for quadrupedal robots in the wild
ANYmal fuses vision and proprioception with zero falls in alpine, forest, underground and urban terrain.
Miki et al. · ETH Zurich, KAIST and IntelScience Robotics · 2020Learning quadrupedal locomotion over challenging terrain
Without vision, ANYmal trots through mud, sand, rubble, thick vegetation, snow and running water.
Lee et al. · ETH Zurich, KAIST and IntelScience Robotics · 2019Learning agile and dynamic motor skills for legged robots
ANYmal runs faster than ever before and recovers from falls, with policies trained in simulation.
Hwangbo et al. · ETH Zurich and IntelCapabilities
Everything your robot needs to learn.
Articulated systemsLegged robots, arms, hands and tracked vehicles, floating or fixed, with any joint type.Contact and frictionHard contacts with Coulomb, rolling and spinning friction, for walking, grasping and pushing.Deformable bodiesCloth, soft solids and deformable terrain that collide with rigid bodies and robots.Granular mediaSand and gravel as individual grains that pile up, flow and push back on robots.Heightmap terrainLarge, rotatable heightmaps, generated procedurally or loaded from images and files.SensorsSpinning LiDAR, RGB and depth cameras and IMUs with noise models, for perception in the loop.Tendons and constraintsSpatial tendons over pulleys, pins, mimic joints and length constraints.Model importURDF, MJCF and OpenUSD robots, plus XML worlds and .rscene scenes.DeterministicThe same inputs reproduce the same trajectory. Save and restore whole worlds for resets and debugging.RaisimServerStream any world to the rayrai viewer over TCP, and pause, step or nudge it from there.Python and RLVectorized C++ environments with PPO in PyTorch, plus RaisimPy for scripting.Cross-platformLinux and Windows on x86-64; Linux and macOS (Apple silicon) on ARM64.
Get going
A running world in a few lines.
- 1Clone raisim2LibThe first CMake configure downloads the matching RaiSim and rayrai binaries.
- 2Build the examplesOne CMake build on Linux, macOS or Windows.
- 3Run and watchStart the TCP viewer and run any server example next to it.
cpp
#include "raisim/RaisimServer.hpp"
#include "raisim/World.hpp"
int main() {
raisim::World world;
world.setTimeStep(0.001);
world.addGround();
auto anymal = world.addArticulatedSystem("rsc/anymal_c/urdf/anymal.urdf");
Eigen::VectorXd q(anymal->getGeneralizedCoordinateDim());
q << 0, 0, 0.54, 1, 0, 0, 0,
0.03, 0.4, -0.8, -0.03, 0.4, -0.8, 0.03, -0.4, 0.8, -0.03, -0.4, 0.8;
anymal->setGeneralizedCoordinate(q);
// PD control on the 12 leg joints, none on the floating base.
Eigen::VectorXd kp = Eigen::VectorXd::Zero(18), kd = Eigen::VectorXd::Zero(18);
kp.tail(12).setConstant(200.0);
kd.tail(12).setConstant(10.0);
anymal->setPdGains(kp, kd);
anymal->setPdTarget(q, Eigen::VectorXd::Zero(18));
// Stream the world to rayrai_tcp_viewer while it runs.
raisim::RaisimServer server(&world);
server.launchServer();
for (int i = 0; i < 100000; ++i) {
RS_TIMED_LOOP(1e6 * world.getTimeStep())
server.integrateWorldThreadSafe();
}
server.killServer();
}Start training today.
Academic and trial licenses are free. Commercial licenses are per person, per year; three or more include a master key that works on any machine.