notesum.ai

Published at October 21

Analyzing Closed-loop Training Techniques for Realistic Traffic Agent Models in Autonomous Highway Driving Simulations

cs.CL

cs.AI

cs.LG

I.2.7; I.2.0

Released Date: October 21, 2024

Authors: Matthias Bitzer, Reinis Cimurs, Benjamin Coors, Johannes Goth, Sebastian Ziesche, Philipp Geiger, Maximilian Naumann

Arxiv: https://arxiv.org/abs/2410.15987v1

Method

Col.

(%)

Off.

(%)

ADE

(m)

Speed

JSD

\times 10^{-2}

Acc.

JSD

\times 10^{-2}

N_{LC}

JSD

\times 10^{-2}

IDM

0.8 ± 0.0

0.00 ± 0.0

5.31 ± 0.0

4.0 ± 0.0

17.64 ± 0.0

0.23 ± 0.0

BC Gaussian-LL

2.67 ± 0.11

6.42 ± 0.42

2.22 ± 0.17

0.63 ± 0.41

8.02 ± 4.6

2.28 ± 0.16

BC GMM-LL

3.1 ± 0.62

7.05 ± 0.78

3.42 ± 0.41

2.41 ± 0.66

23.16 ± 6.52

2.23 ± 0.33

BC wMSE + Orientation

3.1 ± 0.65

4.36 ± 0.73

4.15 ± 0.34

3.51 ± 0.8

12.54 ± 4.33

1.5 ± 0.22

DiffSim MSE

1.1 ± 0.74

3.05 ± 2.11

1.62 ± 0.77

0.28 ± 0.3

1.82 ± 1.57

0.92 ± 1.03

DiffSim wMSE

0.51 ± 0.11

0.68 ± 0.31

1.25 ± 0.13

0.2 ± 0.09

1.05 ± 0.38

0.48 ± 0.09

DiffSim wMSE + Collision

0.39 ± 0.42

2.52 ± 1.15

3.57 ± 3.31

3.24 ± 5.78

2.5 ± 1.7

0.66 ± 0.4

MGAIL + BC-LL (Gauss.)

1.7 ± 0.58

2.95 ± 1.2

1.95 ± 0.24

0.42 ± 0.18

1.82 ± 0.52

0.6 ± 0.56

MGAIL + BC-LL (GMM)

2.87 ± 2.42

2.77 ± 0.52

7.8 ± 8.31

9.34 ± 10.7

18.55 ± 10.69

0.19 ± 0.13

MGAIL + DiffSim wMSE (Gauss.)

0.39 ± 0.06

0.59 ± 0.35

0.93 ± 0.03

0.08 ± 0.03

0.86 ± 0.15

0.32 ± 0.15

MGAIL + DiffSim wMSE (GMM)

0.35 ± 0.19

0.36 ± 0.32

1.0 ± 0.18

0.11 ± 0.1

1.03 ± 0.71

0.31 ± 0.15

MGAIL + Diff. wMSE + Col. (Gauss.)

0.17 ± 0.09

2.28 ± 0.84

1.47 ± 0.09

0.23 ± 0.05

2.35 ± 0.48

0.26 ± 0.21

MGAIL + Diff. wMSE + Col. (GMM)

0.17 ± 0.11

1.49 ± 0.42

1.48 ± 0.16

0.26 ± 0.06

1.86 ± 0.85

0.35 ± 0.42