Human-in-the-loop reinforcement learning (HiL-RL) for robotic foundation models is fundamentally constrained by expensive and risky physical interaction. This makes sample efficiency, stable policy improvement, and reduced human effort essential for practical deployment.

However, existing approaches often suffer from redundant data sampling, biased credit assignment across mixed-quality trajectories, and intervention-intensive exploration.

In this sharing, we explore a unified research direction for improving HiL-RL through informative sample scheduling, credit-aware policy optimization, and agentic intervention. The work covers:

  • Entropy-guided sampling to reduce redundant real-world interaction and improve effective data utilization.

  • Structured credit reassignment to correct overestimated learning targets and stabilize actor-critic updates.

  • Value-aware agentic recovery mechanisms to reduce human intervention while preserving exploration efficiency.

Together, these approaches aim to help HiL-RL systems learn faster, require less human effort, generalize across a broader range of manipulation states, and operate more reliably in complex physical environments.

(Technical Level: 200)

More About the Speaker

Ziwei Wang is currently an assistant professor in School of Electrical and Electronic Engineering, Nanyang Technological University. Before joining NTU, he was a postdoc fellow in Robotics Institute, Carnegie Mellon University.

He received the Ph.D and the B.S degrees from the Department of Automation, Tsinghua University in 2023 and the Department of Physics, Tsinghua University in 2018 respectively.

His research mainly focuses on robotic foundation models for precise manipulation and has been deployed in manufacturing scenarios including electronics assembly and engine inspection.

He has published over 60 scientific papers in top-tier conferences and journals of robotics and AI. He serves as the editorial board member and conference chairs and co-organized many workshops, tutorials and challenges in top journals and conferences. He is a committee member of IEEE RAS Singapore Chapter.

More About the Series

Singapore Embodied AI Colloquium (SEAIC) is a focused technical research series built around researcher-driven discussions and sharing on frontier topics in embodied AI, robot learning, and physical AI.

Get involved: Learn more about Lorong AI | Speaker Sign-up | WhatsApp Community | LinkedIn | X