강화학습을이용한 클라이밍 모션 합성
Climbing Motion Synthesis using Reinforcement Learning
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
최근자연스러운모션데이터에대한수요가늘고있지만, 클라이밍모션을정확하게캡처하는것은가려진부분이많은클라이밍 동작의 특성상 쉽지 않다. 또한 벽 구조물의 스캔이나 다양한 암벽 코스 준비 등 필요한 데이터를 수집하는 과정이 쉽지않다. 본 논문에서는 강화학습을이용한 클라이밍 모션 합성 방법론을 제안한다. 학습 과정은 두 단계의 난이도로 구성되어있다. 첫 번째 단계는 매달리기 정책을 학습하는 것이다. 매달리기 정책은자연스러운 자세로 홀드를 잡는 방법을 학습한다. 이후 추론 단계를 통해 위치, 자세, 잡기 상태를 다양하게 추출한 초기 상태 데이터세트를만든다. 두 번째 단계에서는 이초기 상태 데이터세트를 사용해서 실제 클라이밍을 수행하는 태스크를 학습한다. 클라이밍 정책은자연스러운 자세로 타겟위치로이동하는방법을학습한다. 실험을통해제안하는방법이클라이밍하기위한좋은자세를효과적으로탐색할수있는것을 보였다.
Abstract
Although there is an increasing demand for capturing various natural motions, collecting climbing motion data is difficult due to technical complexities, related to obscured markers. Additionally, scanning climbing structures and preparing diverse routes further complicate the collection of necessary data. To tackle this challenge, this paper proposes a climbing motion synthesis using reinforcement learning. The method comprises two learning stages. Firstly, the hanging policy is trained to grasp holds in a natural posture. Once the policy is obtained, it is used to extract the positions of the holds, postures, and gripping states, thus forming a dataset of favorable initial poses. Subsequently, the climbing policy is trained to execute actual climbing maneuvers using this initial state dataset. The climbing policy allows the character to move to the target location using limbs more evenly in a natural posture. Experiments have shown that the proposed method can effectively explore the space of good postures for climbing and use limbs more evenly. Experimental results demonstrate the effectiveness of the proposed method in exploring optimal climbing postures and promoting balanced limb utilization.
참고문헌 (16)
- S. Maheshwari, "Trans-fer4d : A framework for frugal motion capture and defor-mation transfer", Proceedings of the IEEE, 2023, 12836~12846
- X. Yi, "Transpose : Real-time 3d hu-man translation and pose estimation with sixi nertial sensors", ACM Transactions on Graphics, 40(4) (2021)
- X. B. Peng, "Deepmimic: Example-guided deep reinforcement learn-ing of physics-based character skills", ACM Trans. Graph, 37(4), pp.143:1~143:14 (2018)
- X. B. Peng, "Amp: Adversarial motion priors for stylized physics-based character control", ACM Trans. Graph, 40(4) (2021)
- X. B. Peng, "Ase: Large-scale reusable adversarial skill embeddings for phys-ically simulated characters", ACM Trans. Graph, 41(4) (2022)
- J. Hoand, "Generative adversarial imitation learn-ing", Advances in Neural Information Processing Systems, 29 (2016)
- C. Tessler, "Calm: Conditional adversarial latent models for directable virtual characters", ACM SIGGRAPH 2023 Con-ference Proceedings (2023), Association for Computing Machinery
- Q. Zhu, "Neural categorical priors for physics-based character control", ACM Trans. Graph, 42(6) (2023)[DOI]
- K. Naderi, "Discovering and synthesizing humanoid climbing movements", ACM Trans. Graph, 36(4) (2017)
- K. Naderi, "Learning phys-ically based humanoid climbing movements", Computer Graphics Forum, 37(8), pp.69~80 (2018)
- K. Naderi, "A rein-forcement learning approach to synthesizing climbing move-ments", 2019 IEEE Conference on Games (CoG), pp.1~7 (2019)
- "mixamo" (2020)[URL]
- M. Yan, "Cimi4d: A large multimodal climbing motion dataset under human-scene interactions", Proceedings of the IEEE, 2023, 12977~12988
- J. Schulman, "Proximal policy optimization algorithms", CoRR (2017)
- V. Makoviychuk, "Isaac gym: High performance gpu-based physics simulation for robot learning", CoRR (2021)