| 武林 祥貴 | M, 2回目発表 | ロボットラーニング | 松原 崇充, | 和田 隆広, | 柴田 一騎, | 鶴峯 義久, | 佐々木 光 |
|
title: Feasibility-Aware Imitation Learning from Human Demonstrations
abstract: Autonomous control of legged robots remains a fundamental challenge in robotics, with applications in search and rescue, delivery, and home assistance. One promising approach is to learn from human demonstrations provided through interfaces. Human demonstrations contain rich, task-relevant behaviors that can guide policy learning without requiring direct robot operation for data collection. However, due to embodiment differences between humans and robots, demonstrations may include motions that are physically infeasible for the robot. Directly learning from such data can destabilize training and reduce performance. In this research, we investigate how robots can robustly learn from human demonstrations that may contain low-feasibility segments. We adopt Generative Adversarial Imitation Learning (GAIL), which trains a policy to imitate expert behavior using a discriminator within an adversarial reinforcement learning framework. GAIL cannot explicitly evaluate physical feasibility, which may lead to misleading reward signals and encourage unsafe or infeasible actions. To address this issue, we propose a framework that automatically detects infeasible demonstration segments and reduces their influence during discriminator training to improve the stability and robustness of imitation learning. language of the presentation: English | |||||||
| 冨澤 聖 | M, 1回目発表 | ソフトウェア設計学 | 飯田 元, | 井上 美智子, | 柏 祐太郎, | Reid Brittany | |
|
title: How are snapshot test results reviewed?
abstract: Software testing plays a crucial role in ensuring software quality. In recent years, UI testing has also been adopted in front-end development. However, since test targets tend to be complex, making it difficult to conduct sufficient testing, snapshot testing, which is relatively easy to implement, has come into use. Snapshot testing detects unexpected changes by examining the differences in behavior before and after program modifications. However, currently, reviewers verify visually whether the detected changes are intended. In this study, we collect JavaScript projects on GitHub that use the snapshot testing provided by the Jest testing framework. We then focus on the intention of changes indicated in each commit to investigate how unintended changes are caused and explore the feasibility of a method to predict and point out such changes. language of the presentation: Japanese 発表題目: スナップショットテストはどのようにレビューされるか 発表概要: ソフトウェアテストはソフトウェアの品質を保証する上で重要な役割を果たす．近年では，フロントエンド開発においてもUIテストが利用されるようになった．しかし，テスト対象が複雑になりやすく十分なテストを行うのが困難なため，比較的実装が簡単なスナップショットテストが用いられるようになりつつある．スナップショットテストはプログラムの変更前後の挙動の差分を調べ，予期せぬ変更が起きているかを検出するが，検出された変更が意図したものであるかはレビュワーが目視で確認しているのが現状である．本研究ではJavaScriptプロジェクトの中でもJESTテストフレームワークによって提供されるスナップショットテストを使用しているプロジェクトをGitHub上で収集する．そして各コミットで示される変更意図に着目し，意図しない変更がどのように引き起こされているか，意図しない変更を予測し指摘する方法の実現可能性について調査する． | |||||||
| 吉本 涼茜 | M, 1回目発表 | ソフトウェア設計学 | 飯田 元, | 井上 美智子, | 柏 祐太郎, | Reid Brittany | |
|
title: Is Self-Admitted Technical Debt Tested? Leveraging Test Coverage to Predict Technical Debt
abstract: Self-Admitted Technical Debt (SATD) refers to technical debt that developers explicitly acknowledge through source code comments. While previous studies have shown that SATD-ridden methods are more prone to bugs and frequent changes, there is a lack of empirical evidence regarding how such methods are actually tested. In this study, we analyze eight open-source Java projects to investigate the relationship between SATD and testing efforts using test coverage as a metric. We first identify distinct testing patterns, such as Extra testing care and General postponement, and examine whether the technical type of SATD influences its likelihood of being tested. Furthermore, we evaluate the feasibility of utilizing test coverage to improve the prediction of SATD presence by integrating it into existing machine-learning-based classification models. Our findings provide new insights into current SATD testing practices and the potential for enhancing automated debt detection. language of the presentation: English | |||||||