Bin Fang
Professor, Beijing University of Posts and Telecommunications; “Top-Tier Talent” Professor
Bin Fang is a Professor and doctoral supervisor at Beijing University of Posts and Telecommunications (BUPT), where he is recognized as a “Top-Tier Talent.” He serves as a Council Member of the Chinese Association for Artificial Intelligence (CAAI), Secretary-General of the CAAI Technical Committee on Cognitive Systems and Information Processing, an Outstanding Member of CAAI, a Standing Committee Member of the CCF Technical Committee on Intelligent Robotics, and an IEEE Senior Member. He previously taught in the Department of Computer Science at Tsinghua University. His research interests include embodied intelligence, dexterous manipulation, and foundation models for robotics. The “Tactile Dexterous Hand” developed by his team was selected for the National Science and Technology Innovation Achievements Exhibition during the 13th Five-Year Plan period. His policy consultation reports on humanoid robots and embodied intelligence were adopted by the General Office of the State Council and received instructions from central government leaders. He has received the First Prize of the Natural Science Award from the Chinese Association of Automation and the Special Prize of the World Robot Competition, among other honors.
Topic
Building a Flywheel for Dexterous Manipulation: Iterating Across Embodiment, Data, and Models
Dexterous manipulation remains a critical bottleneck in bringing embodied intelligence from demonstrations to practical applications. A common challenge today is that embodiment hardware, interaction data, and manipulation models are evolving in isolation: hardware lacks sufficient sensing capabilities, data lacks scale, and models lack physical validation. This talk proposes a flywheel approach centered on the iterative development of embodiment, data, and models for dexterous manipulation. Starting with dexterous hands equipped with vision and tactile sensing, high-quality multimodal physical interaction data can be continuously generated through real-world deployment. This data is then fed back into manipulation model training to improve the generalization of grasping and manipulation capabilities. More capable models, in turn, enable the embodiment to operate efficiently across a broader range of scenarios, forming a self-reinforcing closed loop. Drawing on the team’s practical experience deploying vision-tactile sensors, dexterous hands, and related systems, the talk will examine the key conditions for initiating this flywheel—including perception-native design, data collection efficiency, and physical verifiability of models—and explore how a data flywheel can accelerate the practical adoption of embodied intelligence at scale.
