Fengshuo Bai

University of Chinese Academy of Sciences, Shanghai Jiao Tong University

Papers

Total Citations

H-Index

About

Fengshuo Bai is a researcher advancing the frontiers of deep reinforcement learning (DRL), with a focus on multi-task learning and adversarial robustness. His most cited work, "PiCor: Multi-Task Deep Reinforcement Learning with Policy Correction" (2023, 10 citations), tackles a fundamental challenge in multi-task DRL: the inefficiency caused by varying learning speeds and negative gradient interference across tasks. By introducing a policy correction mechanism, PiCor enables more stable and efficient training of generalist agents, offering a practical solution for scaling DRL to complex, multi-objective environments. In his more recent work, "RAT: Adversarial Attacks on Deep Reinforcement Agents for Targeted Behaviors" (2025, 4 citations), Bai shifts focus to security, designing attacks that manipulate DRL agents into specific, attacker-defined behaviors—bypassing traditional reward-based defenses. This research is critical for evaluating and improving the robustness of DRL systems deployed in safety-sensitive applications. Through these contributions, Bai is shaping the future of reliable and versatile reinforcement learning, making his work essential reading for students and researchers interested in both the capabilities and vulnerabilities of modern AI agents.