BayesRace

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

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},
}

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.

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

cd bayes_race/raceline
python generate_raceline_ethz.py

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
```
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
```
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
```
Validate the trained models on the track we want to race.
```
cd bayes_race/gp
python plot_uncertainty_orca.py
```
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.
Benchmark the performance in Step 6 against MPC with true model dynamics.
```
cd bayes_race/mpc
python run_nmpc_orca_true.py
```
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
```

junzengx14/bayesrace

BayesRace

Installation

How to run