Runze Zhang 『张润泽』

I received my Bachelor’s degree of Computer Science and Technology with honors from Tongji University, advised by Prof. Guang Chen. During my undergraduate years, I was fortunate to work with Prof. Auke Ijspeert at EPFL and Prof. Hao Su at UCSD.

My research journey began with classical 2D computer vision, where, as is well known, task-related numerical metrics are the primary criteria for evaluating models. However, I gradually realized that an improved metric does not always yield a “better” vision model in practice; that is, a model’s output is not the final destination. Instead, individual models should serve as vital components within a broader functional pipeline, where their true value is defined by their contribution to the overall system-level objectives. This realization led me to pivot toward robotics, a field that requires integrating various modules, from perception to control, to achieve functional goals such as task success rates in manipulation.

Furthermore, I believe the ultimate goal for most of the leading brains in robotics is to build general-purpose robotic systems that function robustly in the complex real world. Yet, the rise of LLMs, the scaling laws, and the dominance of industry giants have prompted me to rethink my role. Identifying as a researcher rather than an engineer, I find myself more drawn to the underlying mechanisms behind the emerging phenomena or questions in robot learning. I want to uncover the fundamental reasons why certain approaches succeed while others fail and how they work, aiming to model these insights, and use them to guide the refinement or redesign of robotic systems. Consequently, my current research interest is focused on interpretability within robot learning.