Hengyan Liu

Papers

Total Citations

H-Index

About

Hengyan Liu is a researcher advancing the frontiers of reinforcement learning, with a focus on sample efficiency and goal-oriented robotic manipulation. Liu’s key contributions lie in developing novel algorithms that enable agents to learn from sparse rewards more effectively. In their highly cited work, “Episodic Self-Imitation Learning with Hindsight” (2020, 16 citations), Liu introduced a groundbreaking self-imitation algorithm featuring a trajectory selection module and an adaptive loss function. This approach significantly accelerates learning by allowing agents to imitate their own past successful episodes, outperforming traditional self-imitation methods. Building on this, Liu’s “Diversity-based Trajectory and Goal Selection with Hindsight Experience Replay” (2021, 11 citations) tackled the limitations of standard Hindsight Experience Replay (HER) in robotic tasks. By incorporating diversity-driven selection of both trajectories and goals, Liu’s method enhances exploration and learning stability in sparse-reward environments. These contributions are particularly impactful for robotic manipulation, where rewards are often minimal. With a total of 27 citations across these two pivotal papers, Liu’s work is shaping the next generation of efficient, self-improving reinforcement learning systems.