ULTRA provides a gym-based environment using SMARTS to tackle intersection navigation, specifically the unprotected left turn.
Key features:
- Customizable scenarios with different levels of difficulty.
- Analysis tools to evaluate traffic designs including low-mid-high densities.
- Tools to support analyzing social-vehicle behaviors.
- Configurable train and test parameters.
- Benchmark results to train and test custom RL algorithms against.
Get started by choosing an option below.
- Flowchart of ULTRA's interface with SMARTS
- Setting up ULTRA
- Train and Evaluate a Baseline Agent
- Create a Custom Agent
- More about ULTRA Agents
- ULTRA supports RLlib
- Multi-agent Experiments in ULTRA
For a longer introduction to ULTRA, including its purpose, concepts, and benchmarks, please see ULTRA: A reinforcement learning generalization benchmark for autonomous driving.
If you use ULTRA in your research, you can use the following citation.
@misc{elsayed20_ultra,
author = {Elsayed, Mohamed and Hassanzadeh, Kimia and Nguyen, Nhat M.
and Alban, Montgomery and Zhu, Xiru and Graves, Daniel and
Luo, Jun},
journal = {Machine Learning for Autonomous Driving Workshop, Neural
Information Processing Systems},
title = {ULTRA: A reinforcement learning generalization benchmark
for autonomous driving},
url = {https://ml4ad.github.io/files/papers2020/ULTRA: A
reinforcement learning generalization benchmark for autonomous
driving.pdf},
year = 2020,
}