Overview

This repo contains a model predictive controller based on the single rigid body model and written in Python. It cames in two flavours: gradient-based via acados or sampling-based via jax. The controller is tested on real robots and is compatible with Mujoco.

Features gradient-based mpc:

• less than 5ms computation on an intel i7-13700H cpu
• optional integrators for model mismatch
• optional smoothing for the ground reaction forces
• optional foothold optimization
• optional real-time iteration or advanced-step real-time iteration
• optional zero moment point/center of mass constraints

Features sampling-based mpc:

• 10000 parallel rollouts in less than 2ms on an nvidia 4050 mobile gpu!
• optional step frequency adaptation for enhancing robustness
• implements different strategies: random sampling, mppi, or cemppi
• different control parametrizations: zero-order, linear splines or cubic splines (see mujoco-mpc)

Experiments with Aliengo
Simulations with Go2

Dependencies

Gradient-based MPC: It uses CasADI to define the model and acados to solve the optimal control problem. Sampling-based MPC: jax for both. The simulation environment is based on Mujoco.

Installation

1. install miniforge (x86_64)

2. create an environment using the file in the folder installation choosing between nvidia and integrated_gpu:

   conda env create -f mamba_environment.yml

3. clone the other submodules:

   git submodule update --init --recursive

4. activate the conda environment

   conda activate quadruped_pympc_env

5. go inside the folder acados and compile it pressing:

   mkdir build
   cd build
   cmake -DACADOS_WITH_QPOASES=ON  -DACADOS_WITH_OSQP=ON ..
   make install -j4
   pip install -e ./../interfaces/acados_template
   

6. inside the file .bashrc, given your path_to_acados, put:

   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:"/path_to_acados/lib"
   export ACADOS_SOURCE_DIR="/path_to_acados"
   

The first time you run the simulation with acados, in the terminal you will be asked to install tera_render. You should accept to proceed.

How to run

1. activate the conda environment

   conda activate quadruped_pympc_env
   

2. go in the folder simulation and press

   python3 simulation_mujoco.py
   

In the file config.py, you can set up the robot, the mpc type (gradient, sampling..), its proprierties (real time iteration, sampling type, foothold optimization..), and other simulation params (reference, gait type..).

Citing this work

If you find the work useful, please consider citing:

@article{turrisi2024sampling,
  title={On the Benefits of GPU Sample-Based Stochastic Predictive Controllers for Legged Locomotion},
  author={Giulio Turrisi, Valerio Modugno, Lorenzo Amatucci, Dimitrios Kanoulas, Claudio Semini},
  booktitle={arXiv preprint arXiv:2403.11383},
  year={2024}
}

Maintainer

This repository is maintained by Giulio Turrisi.