Sim-to-Real Robot Policies Without the Rewrite

What it does

rl_sar is a C++ framework that ingests policies trained in IsaacGym or IsaacSim and runs them against either physics simulators or physical actuators. It targets quadruped, wheeled, and humanoid platforms from vendors like Unitree, Agibot, FFTAI, and others, handling inference via libtorch or onnxruntime and communication through ROS Noetic or ROS2 Foxy/Humble. The same policy binaries can target Gazebo, MuJoCo, or a real Ethernet-connected robot.

The interesting bit

Instead of treating simulation and hardware as separate projects, rl_sar collapses them into one compatibility layer. The repository even ships pre-trained dance policies—yes, including Gangnam Style for the Unitree G1—and offers an experimental mobile web interface so you can drive supported robots from a phone browser.

Key highlights

• Hardware support spans Unitree A1, Go2, Go2W, G1, and several other models, though some remain simulation-only
• Inference backends include both libtorch and onnxruntime
• Compatible with ROS Noetic and ROS2 Foxy/Humble; macOS is experimental and MuJoCo-only
• Supports both locomotion and dance tasks
• Docker support is available for containerized simulation and deployment

Caveats

• The Python version is temporarily suspended; the README points users to release v2.3 with no committed timeline for re-release
• MuJoCo support is partial, and several robots in the compatibility matrix remain simulation-only or untested on physical hardware
• The author explicitly disclaims liability, warning that wireless robot connections can cause packet loss or loss of control

Verdict

If you have a pile of IsaacGym policies and a legged robot gathering dust in the lab, this is the missing adapter. Everyone else is looking at a very specific piece of robotics glue.