Zicai Peng
Papers
3
Total Citations
8
H-Index
2
About
Zicai Peng is at the forefront of robotic skill acquisition, pioneering methods that enable robots to learn complex behaviors from abundant, low-cost human video data rather than expensive robot-specific demonstrations. His work addresses one of robotics' most pressing bottlenecks: the scarcity of action-labeled data. Peng’s key contributions include **GraphMimic**, a novel graph-to-graphs generative model that translates video observations into diverse, structured policies for robotic learning, and **FMimic**, which harnesses the reasoning power of foundation models—specifically vision-language models (VLMs)—to extract fine-grained action knowledge from human videos. These works, already garnering early citations, promise to dramatically lower the barrier for robotic skill transfer. Beyond video-based learning, Peng also tackles high-precision manipulation challenges, developing visual-tactile fusion techniques for accurate object pose estimation during dynamic interactions like grasping and assembly. By integrating tactile feedback with visual data, his research enhances robotic dexterity in contact-rich tasks. Peng’s innovative, data-efficient approaches are paving the way toward more generalist, adaptable robots capable of learning from the vast repository of human activity.
Research Focus
Key Achievements
Top Papers
- 1
- 2
- 3