optimizers-mpc-jax

Sampling- and gradient-based optimization algorithms in JAX for MPC trajectory planning.

Implemented Optimizers

Sampling-Based

• Cross Entropy Method (CEM)
• Basic Random Search (BRS)
• Augmented Random Search (ARS)
• Stochastic Trajectory Optimization for Motion Planning (STOMP)
• Model Predictive Path Integral (MPPI)

Gradient-Based / Second-Order

• Vanilla Gradient Descent
• Nesterov Accelerated Gradient
• Gauss–Newton Method
• Newton’s Method

Results

Cross Entropy Method (CEM)

Basic Random Search (BRS)

Vanilla Gradient Descent

Nesterov Accelerated Gradient

Gauss-Newton Method

Newton’s Method

Installation

Note that this code has only been tested with ROS-Noetic on Ubuntu 20.04. It is assumed that you already have a complete ROS installation. Also, ensure that CVXOPT is installed.

1. (optional) Set up a catkin workspace

   cd
   mkdir -p catkin_ws/src
   cd catkin_ws
   catkin init
   

2. Clone this repository in the workspace

   cd catkin_ws/src
   git clone https://github.com/Arcane-01/ .
   

3. Build the workspace

   cd ../
   catkin build
   source devel/setup.bash
   

Usage

1. To start the Gazebo simulation with the Jackal and select a specific optimizer, run:

   roslaunch planner jackal_custom.launch world_name:="env4" optimizer:="cem" type:="vanilla"

• world_name - Gazebo world file from planner/worlds. Default is env4.

• optimizer - Optimization algorithm to use in the planner. Options include:

  • cem - Cross Entropy Method
  • random_search - Basic Random Search
  • gauss_newton - Gauss-Newton
  • gradient_descent - Gradient-based methods

• type - Subtype of optimizer (used for gradient-based optimizers). Options include:

  • vanilla - Vanilla Gradient Descent
  • nesterov - Nesterov Accelerated Gradient
  • newton - Newton’s Method

3. Once the planner is running, you can begin navigation by setting a 2D Nav Goal in RViZ.