MC00614 / mppi_ipddp_cpp

C++ Implementation of MPPI-IPDDP (Model Predictive Path Integral - Interior Point Differential Dynamic Programming) and Testing with MPPI variants and Multiple Environments

MPPI-IPDDP C++

Implementation of the MPPI-IPDDP in C++.

Key Features

• Header only library solver
• CPU multi-processing
• Benchmark with MPPI variants and multiple environments
• Collision Checker can be easily modified on demand
• Easy custom model configuration

Usage

Download Dependency

Build Dependancy

EigenRand and autodiff need to be named as in CMakeLists.txt.

• Eigen 3.3.9
• EigenRand
• autodiff

Runtime Only Dependancy

For testing in diffent environment, we use .npy maps from BARN Dataset.

• BARN_dataset

Build with CMake

Note that EigenRand and autodiff need to be under mppi_ipddp_cpp.

git clone https://github.com/MC00614/mppi_ipddp_cpp.git
cd mppi_ipddp_cpp
mkdir build && cd build
cmake.. && make

For simple usage, refer example.cpp with Wheeled Mobile Robot in 2D environment.

Performance comparison with MPPI variants

Box Plot

Seconds	Curvature

Results for MPPI

MPPI	Log-MPPI
Smooth-MPPI	MPPI-IPDDP

To reproduce this, refer main.cpp with Specific Target and Number of Simulations

// Target: MPPI-IPDDP, Simulations: 10
./main MPPI-IPDDP 10

Graphical tool for plot can be found in graph (box.py, graph.py), which is written in Python.

Performance in Multiple Environments (BARN Dataset)

Map Conversion

(a) Original	(b) Extend
(c) Inflate	(d) Distance

Results

Metric	MPPI	Log MPPI	Smooth MPPI	MPPI IPDDP
$N_u$	3200	3200	12800	1600
$𝜮𝒖$	0.2	0.1	0.3	0.4
Success	97	97	91.3	95.7
Q1 Time	0.121522	0.169059	0.446100	0.274369
Q2 Time	0.139242	0.197344	0.511798	0.299426
Q3 Time	0.179549	0.232929	0.598198	0.346971
MSC	0.002528	0.003089	0.005856	0.000139

To test with 300 maps, use test_with_map.cpp with Specific Target

// Target: MPPI-IPDDP
./test_with_map MPPI-IPDDP

Future Works

• GPU Acceleration
• More general constraint handling in model
• Custom implementation of diffential with dual number