Model-Based Reinforcement-Imitation Learning with Mixture-of-Codebooks for Autonomous Driving Simulation
- Gradient Highway + Inter-Agent Pathways for efficient credit assignment
- Dual Regularisation: open-loop IL stabilisation & model-based RL safety constraints
- Mixture-of-Codebooks with temporal abstraction to capture multi-modal behaviour
- State-of-the-art performance on the large-scale Waymo Open Motion Dataset
Note: Full source code, trained models and experiment scripts will be released once the paper is formally accepted.
mric/
ββ docs/
β ββ figs/ # architecture.png
β ββ gifs/ # demo_lane_change.gif, β¦
ββ mric/ # core library (to be released)
ββ examples/ # training / evaluation scripts
ββ README.md
# Clone with submodules if any
git clone --recursive https://github.com/your-org/mric.git
cd mric
# (Optional) Create environment
conda create -n mric python=3.10
conda activate mric
# Install dependencies
pip install -r requirements.txt# 1. Download Waymo Open Motion Dataset (WOMD) and set DATA_ROOT
export DATA_ROOT=/path/to/womd
# 2. Train MRIC
python examples/train_mric.py --config configs/mric_womd.yaml
# 3. Evaluate
python examples/eval_mric.py --checkpoint checkpoints/mric_best.pth| Scenario | GIF Path |
|---|---|
| Highway lane-change | docs/gifs/demo_lane_change.gif |
| Urban intersection | docs/gifs/demo_intersection.gif |
| Emergent behaviour | docs/gifs/demo_emergent.gif |
The high-resolution framework diagram lives at docs/figs/architecture.png and is displayed above.
@article{he2024mric,
title = {MRIC: Model-Based Reinforcement-Imitation Learning with Mixture-of-Codebooks for Autonomous Driving Simulation},
author = {Baotian He and Yibing Li},
journal = {arXiv preprint arXiv:2404.18464},
year = {2024}
}We welcome pull requests once the codebase is public.
Before opening an issue, please search existing ones and read CONTRIBUTING.md.
For questions, please open a GitHub issue or email
Baotian He Β· hbt18@mails.tsinghua.edu.cn
Yibing Li Β· liyb@tsinghua.edu.cn
Β© 2024-present The MRIC Authors. Released under the MIT License.



