RAMAC: Multimodal Risk-Aware Offline Reinforcement Learning and the Role of Behavior Regularization
ICML 2026
Kai Fukazawa
Hi!, I am a Ph.D. Candidate in Mechanical Engineering at the University of California, Davis, advised by Iman Soltani at the Laboratory for AI, Robotics and Automation (LARA).
My current research question is "How can agents make better use of past experience to improve their decision-making while keeping the learning process simple?" To address this, I focus on extracting useful structure from offline data and using it to design simple objectives for stable, safe, and scalable policy learning.
This perspective has led me to work across reinforcement learning and imitation learning for embodied agents, with applications such as robotics.
News
- Passed the Qualifying Examination!
- RAMAC has been accepted to ICML 2026! See you in Seoul, South Korea ✨
Publications
Gaze on the Prize: Shaping Visual Attention with Return-Guided Contrastive Learning
arXiv 2025
Current Projects
Mitigating Compounding Errors in Imitation Learning
We are developing an approach to reduce compounding errors in imitation learning. Our approach leverages additional information from offline demonstrations to improve policy performance without substantially complicating the learning pipeline. Our preliminary experiments show consistent improvements of 10–15% over standard imitation learning baselines, including diffusion policies. This information also provides additional signals for evaluating policy behavior beyond task success, including indicators of distribution shift. We are also working toward real-world bimanual manipulation with ALOHA and building an end-to-end pipeline for teleoperation, data collection, policy training, and evaluation.
Experience
Education
University of California, Davis
Ph.D. in Mechanical and Aerospace Engineering
Kogakuin University
B.E. in Mechanical Engineering
Favourite Quotes
These are a few quotes that have shaped how I think about life and continue to influence my research interests.
- Marcus Aurelius“What stands in the way becomes the way.”
- 題烏江亭 杜牧“勝敗兵家事不期 包羞忍恥是男児 江東子弟多才俊 巻土重来未可知”
- マイク•タイソン“鍛錬を持続しながら臨むことだ。それができなければ、どんなに才能があっても意味がない。”