Zhiheng Wu

University of Chinese Academy of Sciences, Chinese Academy of Sciences

Papers

Total Citations

H-Index

About

Zhiheng Wu is a researcher at the forefront of deep reinforcement learning and underwater computer vision. His work addresses two critical challenges: enabling generalist AI agents through multi-task learning, and overcoming data scarcity in underwater robotics. In his highly cited 2023 paper "PiCor: Multi-Task Deep Reinforcement Learning with Policy Correction," Wu tackles the fundamental problem of negative gradient interference when training a single agent across multiple tasks with varying learning speeds. This work, garnering 10 citations, proposes a policy correction mechanism that significantly improves learning efficiency. Simultaneously, Wu has made substantial contributions to underwater vision, authoring two influential papers on self-supervised image generation. His 2024 paper on underwater domain pre-training and his 2021 work on pixel-level self-supervised synthesis—together accumulating 15 citations—introduce novel methods to generate realistic underwater training data, enabling deep learning models to perform effectively in aquatic environments despite limited real-world datasets. Wu's dual focus on algorithmic efficiency and practical data solutions positions him as a rising innovator bridging reinforcement learning and marine robotics.