Skip to main content

​
Introduction

rfx provides multiple simulation backends that allow you to develop, test, and train robot policies without physical hardware. All simulation backends implement the same unified SimRobot interface, making it easy to switch between simulation and real hardware.

​
When to Use Simulation

Development

Test control logic and algorithms locally without hardware access

Training

Train RL policies with thousands of parallel environments on GPU

Testing

Unit test your robot code with MockRobot (no dependencies required)

Validation

Verify policy behavior before deploying to real hardware

​
Quickstart

The fastest way to get started is with the Genesis backend: 
uv pip install --python 3.13 torch
uv run --python 3.13 rfx/examples/genesis_viewer.py --auto-install
Genesis currently requires Python 3.13 or earlier. Python 3.14 support is pending upstream wheel releases.

​
Minimal Python API

All simulation backends use the same interface: 
from rfx.sim import SimRobot

# Create a simulated robot
robot = SimRobot.from_config(
    "rfx/configs/so101.yaml",
    backend="genesis",      # or "mjx" or "mock"
    num_envs=1,
    viewer=True,
    auto_install=True,
)

# Use the same API as real robots
obs = robot.reset()
action = torch.zeros(1, robot.max_action_dim)
robot.act(action)
obs = robot.observe()

​
Available Backends

rfx supports three simulation backends:
Fast parallel simulation on GPU with built-in visualization.Install: uv pip install genesis-worldBest for: Parallel RL training, rapid prototyping
JAX-accelerated MuJoCo with JIT compilation support.Install: pip install mujoco mujoco-mjx jax[cuda]Best for: Advanced RL research, custom gradients
Lightweight CPU backend with no external dependencies.Install: Built-in (no extra dependencies)Best for: Unit testing, CI/CD, quick experiments

​
Simulation vs Real Hardware

The unified API makes switching between simulation and real hardware seamless: 
from rfx.sim import SimRobot

robot = SimRobot.from_config(
    "rfx/configs/so101.yaml",
    backend="genesis",
    num_envs=4096,  # Parallel environments
)
Both use identical methods:
  • robot.observe() - Get current state
  • robot.act(action) - Send control commands
  • robot.reset() - Reset to initial state

​
Common Configuration

Control simulation behavior with optional parameters: 
robot = SimRobot.from_config(
    "rfx/configs/so101.yaml",
    backend="genesis",
    num_envs=1024,          # Number of parallel environments
    viewer=True,            # Enable 3D visualization
    device="cuda",          # Device placement
    dt=0.002,               # Timestep override
    substeps=4,             # Physics substeps per control step
    auto_install=True,      # Auto-install backend if missing
)

​
Next Steps

Backend Details

Learn about Genesis, MuJoCo MJX, and Mock backends

MockRobot for Testing

Use MockRobot for unit tests without dependencies

Build docs developers (and LLMs) love

Get started for freeTalk to us