A Gaussian Process Model for Opponent Prediction in Autonomous Racing

This repository contains the code for the paper:

A Gaussian Process Model for Opponent Prediction in Autonomous Racing
Finn Lukas Busch, Jake Johnson, Edward L. Zhu, and Francesco Borrelli

Software Requirements

FORCES PRO version 4.9+

Installation instructions

Run install.sh to install barcgp python package.

Data Generation

Run scripts/gen_training_data.py to generate a series of training samples across different track types. This will generate new FORCES controllers to match those used for data generation.

Parameters:

• policy_name: Determines policy that the target vehicle will use in training
• track_types: Track types that will be generated
• total_runs: Number of sample runs

Running Simulations

Run scripts/run_sim.py to simulate a head-to-head race with a predictor modeling the prediction of the target vehicle.

Parameters:

• predictor_class: Predictor to use (type of [ConstantVelocityPredictor, ConstantAngularVelocityPredictor, GPPredictor, NLMPCPredictor])
• policy_name: Determines policy that the target vehicle will use in simulation
• use_GPU: Whether to use GPU for inference when using GPPredictor
• M: Number of samples to generate from GP predictor
• T: Max length in seconds of experiment
• gen_scenario: Controls whether to generate new scenario or use saved pkl
• use_predictions_from_module: Set to true to use predictions generated from predictor_class, otherwise use true predictions from MPCC