Real-Time Neural MPC

This repository contains the code for experiments associated to our paper

Real-time Neural-MPC: Deep Learning Model Predictive Control for Quadrotors and Agile Robotic Platforms

Arxiv Link

If you are looking for the ML-CasADi framework code you can find it here.

Installation

Checkout Submodules

git submodule update --init --recursive

Acados

Follow the installation instructions.
Install the Python Interface.
Ensure that LD_LIBRARY_PATH is set correctly (DYLD_LIBRARY_PATHon MacOS).
Ensure that ACADOS_SOURCE_DIR is set correctly.

Further Requirements

pip install -r requirements.txt

Make sure the ML-CasADi framework is part of the python path.

export PYTHONPATH="${PYTHONPATH}:<path-to-git>/ml-casadi"

Python 3.9 is recommended.

Experiments

The provided code is based on the work of Torrente et al. All functionality of the original code base is retained.

Change the working directory to

cd ros_dd_mpc

Simulation

Data Collection

Run the following script to collect a few minutes of flight samples

python src/experiments/point_tracking_and_record.py --recording --dataset_name simplified_sim_dataset --simulation_time 300

Fitting a MLP Model

Edit the following variables of configuration file in config/configuration_parameters.py (class ModelFitConfig) so that the training script is referenced to the desired dataset. For redundancy, in order to identify the correct data file, we require to specify both the name of the dataset as well as the parameters used while acquiring the data. In other words, you must input the simulator options used while running the previous python script. If you did not modify these variables earlier, you don't need to change anything this time as the default setting will work:

    # ## Dataset loading ## #
    ds_name = "simplified_sim_dataset"
    ds_metadata = {
        "noisy": True,
        "drag": True,
        "payload": False,
        "motor_noise": True
    }

The following command will train an MLP model with 4 hidden layers 64 neurons each to model the residual error on the velocities in x, y, and z direction (7, 8, 9 in the state). We assign a name to the model for future referencing, e.g.: simple_sim_mlp

python src/model_fitting/mlp_fitting.py --model_name simple_sim_mlp --hidden_size 64 --hidden_layers 4 --x 7 8 9 --y 7 8 9 --epochs 100

The model will be saved under the directory ros_dd_mpc/results/model_fitting/<git_hash>/

Fitting GP and RDRv

For instructions on how to fit a GP or RDRv model for comparison see the here

Test the Fitted Model

python src/experiments/trajectory_test.py --model_version <git_hash> --model_name simple_sim_mlp --model_typ mlp_approx

where the model_type argument can be one of mlp_approx(Real-time Neural MPC), mlp (Naive Integration), gp (Gaussian Process Model).

For a baseline comparison result run the same script without model parameters:

python src/experiments/trajectory_test.py

Multiple models can be compared at once via

python src/experiments/comparative_experiment.py --model_version <git_hash_1 git_hash_2 ...> --model_name <name_1 name_2 ...> --model_type <type_1 type_2> --fast

Results are saved in the results/ folder.

Citing

If you use this code in an academic context, please cite the following publication:

@article{salzmann2023neural,
  title={Real-time Neural-MPC: Deep Learning Model Predictive Control for Quadrotors and Agile Robotic Platforms},
  author={Salzmann, Tim and Kaufmann, Elia and Arrizabalaga, Jon and Pavone, Marco and Scaramuzza, Davide and Ryll, Markus},
  journal={IEEE Robotics and Automation Letters},
  doi={10.1109/LRA.2023.3246839},
  year={2023}
}

TUM-AAS/neural-mpc