Eurographics 2021: Learning and Exploring Motor Skills with Spacetime Bounds

1. Introduction

This repository contains code reproduces results in "Learning and Exploring Motor Skills with Spacetime Bounds"(https://milkpku.github.io/project/spacetime.html).

It also contains code implements feature extraction algorithms proposed by "Towards Robust Direction Invariance in Character Animation"(https://milkpku.github.io/project/hairyball.html)

2. Requirements

C++

name	version
Clang	10.0.1 (required)
PyBullet	2.89 (required)
Eigen	3.3.7 (or later)
swig	4.0.2 (or later)
CMake	3.11.0 (or later)
gtest (optional)

In our experiments, we use Clang to compile C++ code and use swig to build modules for Python.

Note:

1. We suspect there are bugs in linux gcc > 9.2 or kernel > 5.3 or our code somehow is not compatible with that. Our code has large numerical errors from unknown source given the new C++ compiler. Please use Clang 10.0.1 or test the project on Windows.
2. Later versions of pybullet remove SPD for joint control, which is critical for our system, thus version 2.89 is required.

Python Pakages:

The version shouldn't matter. You should be safe to install the latest versions of these packages.

name	version
PyTorch	1.8.0 (or later)
numpy	1.19.1 (or later)
numba	0.50.0 (or later)
tensorboardX	2.2 (or later)

3. Setup

Before running pretrained models or training new models, code in folder Kinematic should be compiled. Following are compile commands:

cd Kinematic && mkdir build && cd build
cmake .. -DEIGEN_INCLUDE_PATH=<path_to_eigen_src_dir> -DPYTHON_INCLUDE_PATH=<path_to_python_include_dir>
make -j8

You should replace <path_to_eigen_src_dir> and <path_to_python_include_dir> with your own path to Eigen source and Python include files.

4. Usage

Since we use multiprocessing in Python, there may be conflicts with OpenMP used by NumPy, so run

export OMP_NUM_THREADS=1

before training models.

To run pretrained models, firstly clone repo from https://github.com/milkpku/spacetimeBounds_policies and copy the contents to path spacetimeBounds/data/policies, then run:

python run_model.py args/demo_spacetime_walk.json

To train models:

python train.py args/train_spacetime_run.json --id spacetime_run