KELab¶

KELab is a reinforcement-learning framework built on KangEngine. It connects KangEngine simulation and rendering to Gymnasium tasks and training workflows for RL Games, RSL-RL, and MimicKit.

KELab is maintained as a separate project. Use its repository for task and training documentation; this page covers the shortest KangEngine setup path.

Install¶

Until KangEngine wheels are available, build KangEngine and install both projects into the KELab environment:

# KangEngine
cd /path/to/KangEngine
make build_python_cuda

# KELab
cd /path/to/KELab
uv venv --python python3.12
source .venv/bin/activate
uv pip install -e /path/to/KangEngine/python
uv pip install -e ".[rl-games]"

Verify the installation:

python -c "import kangengine, ke_lab, ke_lab_tasks, ke_lab_rl"

Run an environment¶

Run a registered task with random actions and KangEngine rendering:

python scripts/random_agent.py \
  --task Humanoid-v0 \
  --num_envs 4 \
  --device cuda \
  --render

For headless training, increase --num_envs, omit --render, and use one of the training scripts under scripts/rl_games, scripts/rsl_rl, or scripts/mimickit.

KELab currently targets PhysX, with its high-throughput training path intended for Linux systems with NVIDIA CUDA. CPU execution remains useful for small smoke tests and debugging.

See the KELab repository for available tasks, training commands, and optional integrations.