andreas122001/temporal_tfpp

Temporal TransFuser++

About

This repo contains the model weights for adapted temporal versions of the TransFuser++ model by Jaeger et. al. 2023.

These models were trained as part of a master's thesis titeled "Temporal Fusion in Imitation-Based End-to-End Autonomous Driving" at the Norwegian University of Science and Technology (NTNU). The thesis will be linked when it becomes available.

All model weights and checkpoint states can be found under "Files and versions".

Results on Bench2Drive

Experiment	Overall		Multi-ability metrics
	DS	SR	Merge	Overtake	Emergency Brake	Give Way	Traffic Sign	Mean
static-8x1	81.68	62.27	57.92	44.44	76.67	46.67	78.42	60.82
static-8x2	83.22	64.24	59.68	45.19	81.11	50.00	79.65	63.13
lidar-4x1	63.58	36.52	33.07	20.00	35.00	46.67	67.54	40.46
full-4x1	79.78	56.06	47.08	41.48	71.67	50.00	76.84	57.41
static-4x1	81.25	60.45	50.63	46.67	80.00	50.00	77.19	60.90
notraj-4x1	81.14	60.45	54.62	45.19	77.22	50.00	74.56	60.32
noprune-4x1	82.80	63.79	54.20	49.63	83.79	50.00	80.00	63.52
large-4x1	82.22	63.03	58.33	40.74	80.00	50.00	81.75	62.17
Default TF++	80.87	62.58	58.33	42.22	74.44	46.67	80.88	60.51

DS = driving score

SR = success rate

Multi-ability metrics are defined by Bench2Drive [ArXiv].

Credits

• CARLA Garage repository by Jaeger et. al.
• Hidden Biases of End-to-End Driving Models by Jeager et. al. 2023
• Hidden Biases of End-to-End Driving Datasets by Zimmerlin et. al. 2024
• Bench2Drive by Jia et. al. 2024