mungi / bayesrace

Learning based control for autonomous racing

BayesRace

BayesRace is a model-based planning and control framework for autonomous racing. It focuses on two problems:

1. Computing the racing line using Bayesian optimization PDF

@InProceedings{JainRaceOpt2020,
  author    = {Jain, Achin and Morari, Manfred},
  title     = {{Computing the racing line using Bayesian optimization}},  
  booktitle = {Proceedings of the 59th IEEE Conference on Decision and Control (CDC)},
  year      = {2020},
}

2. Designing a self-learning controller that reduces the effort required for system identification by learning from prior experience PDF

@InProceedings{JainBayesRace2020,
  author    = {Jain, Achin and O'Kelly, Matthew and Chaudhari, Pratik and Morari, Manfred},
  title     = {{BayesRace: Learning to race autonomously using prior experience}},  
  booktitle = {Proceedings of the 4th Conference on Robot Learning (CoRL)},
  year      = {2020},
}

Installation

We recommend creating a new conda environment:

conda create --name bayesrace python=3.6
conda activate bayesrace

Then install BayesRace:

git clone https://github.com/jainachin/bayesrace.git
cd bayesrace/
pip install -e .

How to run

The following steps are explained for the 1:43 scale autonomous racing platform at ETH Zurich. We also provide code for the 1:10 scale F1TENTH racing platform at University of Pennsylvania.

1. Compute the racing line for the track we want to race on.

   cd bayes_race/raceline
   python generate_raceline_ethz.py
   

   

2. Run a pure pursuit controller on a different track to log sensor measurements and state estimates. This data resemble true system dynamics.

   cd bayes_race/pp
   python run_pp_orca.py
   

   

3. Given a trajectory of states from Step 2, generate an equivalent trajectory using a simpler and easy-to-tune e-kinematic model.

   cd bayes_race/data
   python simulate_kinematic_orca.py
   

4. Train Gaussian process models to predict mismatch between true system dynamics from Step 2 and e-kinemtic model from Step 3.

   cd bayes_race/gp
   python train_model_orca.py
   

5. Validate the trained models on the track we want to race.

   cd bayes_race/gp
   python plot_uncertainty_orca.py
   

   

6. Run MPC with and without GP correction. By default boundary constraints are turned off for faster execution.

   cd bayes_race/mpc
   python run_nmpc_orca_gp.py
   

   MPC without GP correction is shown on the left, and with GP correction on the right.

   

7. Benchmark the performance in Step 6 against MPC with true model dynamics.

   cd bayes_race/mpc
   python run_nmpc_orca_true.py
   

   

8. Finally, update the GP models using data collected in Step 6 that is specific to the racing track and re-run MPC with the updated GP models.

   cd bayes_race/gp
   python update_model_orca.py
   

   cd bayes_race/mpc
   python run_nmpc_orca_gp_updated.py