CIRL: Controllable Imitative Reinforcement Learning for Vision-based Self-driving

This repo is a inofficial tensorflow implemention of ECCV2018 Paper：CIRL: Controllable Imitative Reinforcement Learning for Vision-based Self-driving. The implementation is to run on CARLA 0.8.4 stable version only.

Different from the original implementation in CIRL paper, we build this codebase to provide a baseline for another paper, IPC: Instance-Aware Predictive Navigation inMulti-Agent Environments. So some components are different, including:

1. the termination condition of each episode is different.
2. the maximum speed is 20m/s instead of 10m/s in original CIRL paper.
3. the target is to drive as safely and fast as possible instead of reaching some preset destination.
4. the evaluation is based on a modified reward.

All difference is to align the implementation with IPC and SPC for fair comparison.

To pretrain the model, we provide a collected on CARLA 0.8.4 with full stack of information, such as action trajectories, depth map, RGB observations, segmentations and so on. Downlaod the dataset here.

To cite the CIRL paper with bibtex.

If you use the codebase together with the agent definition borrowed from IPC or SPC, please cite them as well: SPC and IPC.