Flexible Conditional Imitation Learning

The purpose of this project is to implement the Conditional Imitation Learning on various scenarios with Airsim Car plugin and Carla Environments. The original Conditional Imitation Learning propose a method in which a single navigatorial command is employed to take a certain action in perplexing situations. Instead, to increase the flexibility, this project concentrates on conditions similar to a path--a set of points on a plane--whereby we could present any kind of turns and velocities.

Requirements

• Tensorflow
• Unreal Engine 4 (UE4)
• Airsim
• Carla

Experiments

Three different strategies are assessed.

Mountainous UE4 Environment

In this environment, the goal is to create an autonumous agent to drive through a tirtuous and mountainous road with the help of Behavioral Cloning and Dataset Aggregation approaches.

Behavorial Cloning

Dataset Aggregation (DAgger)

Collecting Data for Dataset Aggregation

Demo

Autonomous Car Driving in Mountainous Environment

Neighborhood UE4 Env

This Env

2022-10-18.23.24.56.mp4

Untitled.mp4

Conditional Imitation Learning

Carla Town

Flexible conditions

Path Prediction

Translator

References

1. Codevilla, F., Müller, M., López, A., Koltun, V., and Dosovitskiy, A., “End-to-end Driving via Conditional Imitation Learning”, arXiv e-prints, 2017.
2. Codevilla, F., Santana, E., Lopez, A., & Gaidon, A. (2019). Exploring the limitations of behavior cloning for autonomous driving. Proceedings of the IEEE International Conference on Computer Vision, 2019-Octob(Cvc), 9328–9337. https://doi.org/10.1109/ICCV.2019.00942
3. Rhinehart, N., McAllister, R., & Levine, S. (2018). Deep Imitative Models for Flexible Inference, Planning, and Control. 1, 1–19. http://arxiv.org/abs/1810.06544