EMTS for Autonomous Driving | 20231005个人简介

Project Description

This project aims to address the long-term planning tasks in autonomous driving missions.
This project will perform trajectory-level search instead of individual control variables. This approach effectively reduces the number of search nodes while enhancing the efficiency of the search process.
This project utilizes expert data of various styles to assist in training Monte Carlo Tree Search (MCTS) and ensure the multimodality of the search strategy in each iteration of MCTS.
The effectiveness of the method has been validated through experimental verification.
More information can be found in paper 1 and paper2 .
This project is still being updated, and you are welcome to check back anytime.

Currently, the AlphaZero and MuZero algorithms are gaining popularity in the field of autonomous driving.
TThe utilization of MCTS (Monte Carlo Tree Search) enables long-term planning, modeling uncertainty, and enhancing the interpretability of driving tasks.
However, they still face challenges. If the planning time is too long, the depth of the tree will increase significantly, leading to an exponential increase in the number of search nodes. Additionally, sample efficiency is also a pressing issue that needs to be addressed.
To overcome the aforementioned challenges, we have proposed two solutions.
- Incorporate motion primitive methods into MCTS : Conduct Searches based on a sequence of motions (skill-level) over a specific duration, rather than indicidual actions at each moment.
- Expert policy involvement search: Retain expert skills as candidate options, but without imposing a mandatory selection of them
The combined application of these two solutions holds the promise of addressing the issues of long planning time, high node count, and low sample efficiency, thereby improving the performance and reliability of autonomous driving tasks.

Long horizon planning demonstration

Trajectory-level MCTS visualization