Scalable Deep Kernel Gaussian Process for Vehicle Dynamics in Autonomous Racing

Paper | Conference

Vehicle Dynamics Modeling Is Challenging​

• Building dynamic vehicle models capable of modeling non-linear behaviors
• Obtaining physics-based model coefficients is often time and cost prohibitive.​
  • Learning drivetrain dynamics requires dyno testing
  • Understanding tire dynamics involves experimenting with specialized tire rigs.
    Slide image

Approximate Dynamics Using Simpler Model

single-track models

• Approximate the observed dynamics with a simpler model (single-track model)
• Build a GP model to capture the residuals between the simpler model output and observations.
  Ekin Model $+$ GP $\approx$ Observed Dynamics
  Slide image

DKL-SKIP GP Method

• DKL captures the most relevant information while reducing its dimensionality.

$$ k(d_i, d_j;\theta) \rightarrow k(g(d_i,w),g(d_j,w)|\theta,w) $$

• SKIP-GP utilizes SKI & product kernel structure to reduce the computing complexity of GP
  Slide image

Experiment Setup

• Conduct experiments using data collected from a real-world autonomous Indy car @ LVMS.
• Collect data from Autoverse, a high-fidelity racing simulator @ TMS.
• Compare the open loop one-step prediction performance of DKL-SKIP with the SKIP-GP and N4SID method.​
  Slide image

Results

• Calculate the MAE, Root mean square error, normalized root mean square error and R2 coefficient for each error term
• DKL-SKIP outperforms the other methods in both real-world and simulated experiments.
  Slide image

Future Work

• Closed-loop MPC implementation
• Multi-output DKL-SKIP model​