Legged RL Deploy

legged_rl_deploy deploys learned policies on top of unitree_lowlevel for Unitree Go2 and G1 robots.

It provides:

• single-policy deployment from a single config.yaml
• multi-policy deployment with gamepad-triggered switching via fsm
• ONNX Runtime and LibTorch inference backends
• local-file and Redis-based mimic motion sources
• sample policies and deployment configs for Go2 and G1

Overview

This package inherits the low-level control loop, safety logic, and hardware adapters from unitree_lowlevel. The RL policy only becomes active after the low-level controller enters HighController.

The deployed config is passed directly to:

ros2 run legged_rl_deploy legged_rl_deploy_node <network-interface> <config-file>

Two config styles are supported:

• Single-policy config: top-level policy: section
• Multi-policy config: top-level fsm: section with multiple sub-policies and gamepad transitions

Video demo:

• https://renkunzhao.github.io/present/slides/phd-interview/#/5

Policy Layout

The repository already includes several example policies:

policies/
├── go2/
│   ├── My_unitree_go2_gym/
│   │   ├── trot/
│   │   ├── hop/
│   │   ├── handstand/
│   │   ├── legstand/
│   │   ├── spring_jump/
│   │   └── multi-policies.yaml
│   ├── mjlab/
│   │   ├── velocity/
│   │   ├── hop/
│   │   ├── mimic/
│   │   └── multi-policies.yaml
│   ├── unitree_rl_lab/
│   │   └── velocity/
│   └── aba/
│       └── velocity/
└── g1/
    ├── velocity/
    │   ├── unitree_rl_mjlab/
    │   └── mjlab/
    ├── mimic/
    │   ├── unitree_rl_lab/
    │   │   ├── dance_102/
    │   │   └── gangnam_style/
    │   ├── unitree_rl_mjlab/
    │   │   └── dance1_subject2/
    │   ├── mjlab/
    │   │   └── dance1_subject1/
    │   ├── whole_body_tracking/
    │   │   └── jumps1_subject1/
    │   └── TWIST2/
    ├── aba/
    │   ├── velocity/
    │   ├── dance102/
    │   ├── jumps1_subject1/
    │   └── multi-policies.yaml
    └── multi-policies.yaml

Included policy sources referenced by this repo:

Source	Link
unitree_rl_lab	github.com/unitreerobotics/unitree_rl_lab
My_unitree_go2_gym	github.com/yusongmin1/My_unitree_go2_gym
mjlab	github.com/mujocolab/mjlab
unitree_rl_mjlab	github.com/mujocolab/unitree_rl_mjlab

Some configs in this repo are experiment-specific derivatives or internally retargeted motions.

Dependencies

Required

• unitree_lowlevel built in the same workspace
• ROS 2 workspace already initialized and buildable with colcon

Recommended for the bundled policies

Most bundled policies use backend: ort, so ONNX Runtime is effectively required for the default examples:

cd $WORKSPACE/src/legged_rl_deploy
./scripts/get_ort.sh

This downloads ONNX Runtime into thirdparty/onnxruntime.

Optional: LibTorch

Use LibTorch only if your policy config sets backend: libtorch or backend: torch.

Helper script:

cd $WORKSPACE/src/legged_rl_deploy
./scripts/get_torch.sh

Platform-specific notes are in LibTorch.md.

Optional: Redis mimic source

Redis support is compiled in only when hiredis is available at build time:

sudo apt install -y libhiredis-dev redis-tools

This is only needed for configs using mimic.params.source: redis, such as policies/g1/mimic/TWIST2/config.yaml.

Build

Clone the repository into the workspace:

cd unitree_ws/src
git clone https://github.com/Renkunzhao/legged_rl_deploy.git

If you plan to use:

• ONNX Runtime: run ./scripts/get_ort.sh first
• LibTorch: install it and set environment variables as described in LibTorch.md before building
• Redis mimic: install libhiredis-dev before building

Then build:

cd unitree_ws
source /opt/ros/<ros-distro>/setup.bash
colcon build --packages-up-to legged_rl_deploy --cmake-args -DPython3_EXECUTABLE=/usr/bin/python3
source install/setup.bash

Note: The explicit Python3_EXECUTABLE avoids clean-build failures when python3 points to a Conda environment without the ROS build dependencies.

Running

Hardware preparation

legged_rl_deploy still relies on the low-level controller setup from unitree_lowlevel.

For Go2 hardware, stop the default controller once after each boot:

source src/unitree_lowlevel/scripts/setup.sh <network-interface> <ros-distro>
$WORKSPACE/build/unitree_sdk2/bin/go2_stand_example <network-interface>

For G1 hardware, enter Debug Mode first as described in the unitree_lowlevel documentation.

Recommended launcher

From the workspace root, use the helper wrapper:

./src/legged_rl_deploy/scripts/run.sh <network-interface> <ros-distro> ros2 run legged_rl_deploy legged_rl_deploy_node <network-interface> <config-file>

The wrapper:

• sources unitree_lowlevel/scripts/setup.sh
• sets CMAKE_PREFIX_PATH and LD_LIBRARY_PATH for Torch on x86_64
• sets the corresponding Torch paths from Python on aarch64

Runtime behavior

The node inherits the same safety confirmation as unitree_lowlevel:

WARNING: Make sure the robot is hung up or lying on the ground.
Press Enter to continue...

In practice, the flow is:

1. Start the node
2. Confirm the safety prompt
3. Use the low-level controller gamepad flow to reach FixStand
4. Press START to enter HighController
5. The RL policy then takes control

If you are using a multi-policy config, policy switching is handled by the fsm.policies.*.transitions entries in the YAML file.

Config Notes

Single-policy configs typically define:

• llc_config_file: low-level config from unitree_lowlevel
• ll_dt: low-level timestep override
• policy.backend: ort, onnxruntime, torch, or libtorch
• policy.model_path: model file path relative to $WORKSPACE
• policy.observations: observation terms and assembly layout
• policy.actions: output post-processing
• policy.commands: command preprocessing

Multi-policy configs additionally define:

• clip_final_tau: optional torque-mode output instead of PD targets
• fsm.default: policy activated when entering HighController
• fsm.policies.<name>.config: sub-policy config path
• fsm.policies.<name>.transitions: button mappings for switching

Teleop and Redis Mimic

For Redis-based mimic motion:

sudo apt install -y libhiredis-dev redis-tools

Rebuild legged_rl_deploy after installing hiredis, otherwise Redis support will remain disabled in the compiled binary.

You can inspect Redis keys with:

redis-cli -h 127.0.0.1 -p 6379 -n 0 --scan

Troubleshooting

backend=onnxruntime requested, but built without USE_ORT

Run:

cd $WORKSPACE/src/legged_rl_deploy
./scripts/get_ort.sh

Then rebuild the package.

backend=libtorch requested, but built without USE_TORCH

Install LibTorch, export the required paths from LibTorch.md, and rebuild the package.

Redis mimic complains that hiredis is missing

Install libhiredis-dev, then rebuild the package so USE_HIREDIS is enabled.

CMake CUDA dialect error

If you see an error like:

Target "cmTC_xxxx" requires the language dialect "CUDA17", but CMake does not know the compile flags to use to enable it.

Update CMake to a newer version that supports CUDA17.