Model-Based Imitation Learning for Urban Driving

Authors: Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, Jamie Shotton

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental MILE improves upon prior state-of-the-art by 31% in driving score on the CARLA simulator when deployed in a completely new town and new weather conditions.
Researcher Affiliation Collaboration Anthony Hu1,2 Gianluca Corrado1 Nicolas Griffiths1 Zak Murez1 Corina Gurau1 Hudson Yeo1 Alex Kendall1 Roberto Cipolla2 Jamie Shotton1 1Wayve, UK. 2University of Cambridge, UK.
Pseudocode No The paper describes its architecture and models using mathematical formulations and textual descriptions, and refers to Appendix C for network details, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes The code and model weights are available at https://github.com/wayveai/mile.
Open Datasets No The training data was collected in the CARLA simulator with an expert reinforcement learning (RL) agent [9] that was trained using privileged information as input... We collect data at 25Hz in four different training towns (Town01, Town03, Town04, Town06) and four weather conditions... for a total of 2.9M frames, or 32 hours of driving data. While the CARLA simulator [8] is publicly available, the specific dataset collected by the authors for this paper is not provided with a direct link or public repository.
Dataset Splits No The paper describes its training data (2.9M frames) and testing environment (Town05), but does not explicitly detail a separate validation set split or its purpose during training.
Hardware Specification Yes Our model was trained for 50, 000 iterations on a batch size of 64 on 8 V100 GPUs
Software Dependencies Yes All experiments were performed on CARLA 0.9.10, an open-source urban driving simulator [8], which comes with pre-built maps [52].
Experiment Setup Yes Our model was trained for 50, 000 iterations on a batch size of 64 on 8 V100 GPUs, with training sequence length T = 12. We used the Adam W optimiser [44] with learning rate 10 4 and weight decay 0.01.