RL With MimicKit

KangEngine can be used as a backend engine of MimicKit through KangEngine’s Python package. Use the backend_kangengine branch of MimicKit and keep MimicKit in a separate Python environment.

Note: GPU contact sensors provide contact count, contact state, and accumulated normal impulse and force. Tangential friction impulse and a full six-axis contact wrench are not currently exposed as sensor outputs.

MimicKit setup and run commands

1. Clone the KangEngine-enabled MimicKit fork branch.

   git clone -b backend_kangengine https://github.com/Gudegi/MimicKit.git
   

2. Create and activate a MimicKit Python environment with uv.

   cd MimicKit
   uv venv .venv --python 3.12
   source .venv/bin/activate
   

3. Build KangEngine’s Python extension from the KangEngine repo.

   cd /path/to/KangEngine
   make build_python
   

4. Install KangEngine’s Python package into the MimicKit environment.

   uv pip install -e ./python
   

5. Install MimicKit dependencies.

   cd /path/to/MimicKit
   uv pip install -r requirements.txt
   

6. Run a small motion visualization test.

   python mimickit/run.py \
     --mode test \
     --num_envs 1 \
     --engine_config data/engines/kangengine_engine.yaml \
     --env_config data/envs/view_motion_humanoid_env.yaml \
     --visualize true \
     --devices cpu \
     --test_episodes 10
   

7. Run pretrained policy inference with KangEngine.

   python mimickit/run.py \
     --mode test \
     --num_envs 4 \
     --engine_config data/engines/kangengine_engine.yaml \
     --env_config data/envs/amp_humanoid_env.yaml \
     --agent_config data/agents/amp_humanoid_agent.yaml \
     --visualize true \
     --model_file data/models/amp_humanoid_spinkick_model.pt
   

8. Train an AMP policy with KangEngine.

   python mimickit/run.py \
     --mode train \
     --num_envs 4096 \
     --engine_config data/engines/kangengine_engine.yaml \
     --env_config data/envs/amp_humanoid_env.yaml \
     --agent_config data/agents/amp_humanoid_agent.yaml \
     --visualize false \
     --out_dir output/
   

For reference, the MimicKit KangEngine backend uses an engine config like this:

engine_name: "kangengine"

control_mode: "pos"
control_freq: 30
sim_freq: 120
env_spacing: 5
enable_self_collisions: false

The backend_kangengine branch already includes data/engines/kangengine_engine.yaml, so you usually do not need to create it manually.