Intro

Status: Archive (code is provided as-is)

Mujoco version of DeepMimic:

• No C++ codes --> pure python
• No bullet engine --> Mujoco engine
• No PPO --> TRPO-based

Examples:

• Walk (play MoCap data):

• Spinkick (play MoCap data):

• Dance_b (play MoCap data):

• Stand up straight (training via TRPO):

Install

• Mujoco: Download mujoco200 and put it in the ~/.mujoco/ folder (mjkey.txt should also be in this folder). Then install mujoco-py:

python3 -m pip install mujoco-py

• python3 modules: python dependencies

python3 -m pip installl gym
python3 -m pip install tensorflow-gpu
python3 -m pip install pyquaternion
python3 -m pip install joblib

• MPI & MPI4PY: mpi for parrellel training

sudo apt-get install openmpi-bin openmpi-common openssh-client libopenmpi-dev
python3 -m pip install mpi4py

Usage

• Testing examples:

python3 dp_env_v3.py # play a mocap
python3 env_torque_test.py # torque control with p-controller

• Gym env

Before training a policy: Modify the step in dp_env_v3.py to set up correct rewards for the task. Use dp_env_v3.py as the training env.

Training a policy that makes the agent stands up straight:

python3 trpo.py

Running a policy:

python3 trpo.py --task evaluate --load_model_path XXXX # for evaluation
# e.g., python3 trpo.py --task evaluate --load_model_path checkpoint_tmp/DeepMimic/trpo-walk-0/DeepMimic/trpo-walk-0

Acknowledge

This repository is based on code accompanying the SIGGRAPH 2018 paper: "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills". The framework uses reinforcement learning to train a simulated humanoid to imitate a variety of motion skills from mocap data. Project page: https://xbpeng.github.io/projects/DeepMimic/index.html