Han‐Bo Zhang

Xi'an Jiaotong University, National University of Singapore

Papers

Total Citations

H-Index

About

Han-Bo Zhang is a leading researcher in robotic manipulation and reinforcement learning, with a focus on enabling robots to perceive, grasp, and interact with objects in unstructured environments. His most impactful work, the fully convolutional grasp detection network with oriented anchor boxes (2018, 22 citations), introduced a real-time, end-to-end approach for predicting multiple grasping poses from RGB images—a foundational contribution to vision-based robotic grasping. Zhang has also made significant advances in tackling the sparse reward problem in reinforcement learning. He developed Hindsight Trust Region Policy Optimization (HTRPO, 2019, 8 citations), which extends the popular TRPO algorithm with hindsight to improve learning efficiency in tasks with minimal feedback. Further extending this line of work, his density-based curriculum for multi-goal RL (2021, 2 citations) provides a principled method for automatically generating training tasks, reducing the need for manual reward engineering. Most recently, his 2025 work on vision-language-action models explores the critical design choices for building generalist robot policies. With a career spanning foundational grasp detection to modern policy learning, Zhang’s research consistently pushes toward more capable, autonomous, and generalizable robotic systems.