ASSET Symposium in conjunction with ACM MobiSys2024 (Pang)

学会情報

会議概要

• Reference URL: https://www.sigmobile.org/mobisys/2024/asset.html

MobiSys 2024 is featuring a special program called ASSET (Asian Student Symposium on Emerging Technologies) as part of SIGMOBILE’s student outreach program. Student participating to ASSET will be divided into small groups and each group will be provided with an excellent and dynamic research mentor who will help them hone their research skills.

MobiSys 2024では、SIGMOBILEの学生アウトリーチプログラムの一環として、ASSET（アジア学生新興技術シンポジウム）という特別プログラムが開催されます。ASSETに参加する学生はグループに分けられ、それぞれのグループには優秀で活気的な研究メンターがいて、彼らの研究スキルを磨く手助けをします。

発表概要

タイトル

ASSET Poster: A Case-based Reward Function Design for Enhanced Performance in Reinforcement Learning-based Pure Pursuit Path Tracking

内容

In autonomous driving, the traditional pure pursuit tracking algorithm is robust in many scenarios but requires manual adjustments. Prior research has shown that combining reinforcement learning with pure pursuit can achieve high tracking accuracy at low speed profile on simple path shape. However, maintaining tracking accuracy while changing speeds is critical but challenging to balance effectively, it requires extensive manual parameter tuning for different speeds. To do so, we improved a reinforcement learning-based pure pursuit controller. We implemented our work in the CARLA simulator. The preliminary experiment shows that by incorporating future curvature into the observation space and reward function, the vehicle agent can learn a faster and more accurate path tracking policy.

自動運転において、従来のPure Pursuit 追跡アルゴリズムは多くのシナリオに対しては頑健であるが、手動での調整が必要だと考えられる。先行研究によると、強化学習とピュアパシュートを組み合わせることで、単純な経路形状において低速で高い追跡精度を達成できることが示されている。しかし、速度を変更しながら追跡精度を維持することは重要である一方で、効果的にバランスを取るのが難しく、異なる速度に対して広範な手動パラメータ調整が必要である。この問題を解決するために、我々は既存の強化学習に基づくピュアパシュートコントローラを改善した。研究はCARLAシミュレータに実装された。予備実験の結果、観測空間と報酬関数に将来の曲率を組み込むことで、車両エージェントがより速く、より正確な経路追跡ポリシーを学習できることが示されている。

質疑

• Q1: Since you are researching path tracking, why don't you try using some optimal-based algorithms such as LQR?

                 経路追跡を研究しているのなら、LQRのような最適ベースのアルゴリズムを試してみてはどうですか？

• Q2: How does the case-based reward function affect the learning?

                 ケースベースの報酬関数は学習にどのように影響しますか？

体験記

This is my first time attending an international conference and having a poster presentation. It was a refreshing experience for me to talk to so many researchers from different countries and different research topics. At the beginning, it was really hard to give the audience a pitch to make them understand my research quickly, but after 2 or 3 rounds, it got better. Also, I had the chance to network with students from different universities, such as Osaka University. It was a joyous moment sharing our research with each other.

The main conference program is filled with interesting research, I have gained insights into various research directions and ways of thinking. I believe this will be very helpful for my future research.

By the way, the banquet was so luxury!

まとめ

This truly was a valuable and interesting experience attending the conference. My English skills were honed, and I made new friends. The journey was exhausting, but I returned with a wealth of knowledge.