Hengyi (Henry) Zhou

I am a junior student with a dual-degree program in Mechanical Engineering at Shanghai Jiao Tong University and Computer Engineering at the University of Michigan. I once worked in SAIL Lab led by Professor Chao Li at SJTU. After arriving at the University of Michigan, I joined the Hybrid Dynamic Robotics Lab led by Professor Xiaonan (Sean) Huang. Also, I collaborated with Professor M. Khalid Jawed closely.

My research interests have gradually transitioned from high-performance computing to robotics. I aim to develop into a full-stack robotics researcher, bringing AI methods into the physical world. Currently, my research focuses on the simulation and manipulation of soft structures and sensors through machine learning methods.

I am actively seeking a Ph.D position in robotics, CS, ECE or related fields starting in Fall 2027.

Email  /  CV (2026.09)  /  Github /  Linkdin

Research

I am committed to leveraging AI-driven approaches to tackle challenges in the optimization and design problems in robotics and to building reliable systems which interacts between intelligence and real-world environments. Ultimately, I aspire for my research to promote technological accessibility and create tangible societal impact. Currently, my research focuses on the simulation and manipulation of soft structures and sensors through machine learning methods.

SCNC learns trajectories for implicit equilibrium systems while reducing gradient errors and avoiding undesirable equilibrium-branch switching by explicitly controlling higher-order sensitivity variations.

Proposing Neural Control, a memory-efficient adjoint-based framework for controlling multi-stable physical systems by computing trajectory-aware gradients without unrolling equilibrium solvers, enabling robust long-horizon manipulation of deformable objects.

Introducing new PCR system that improves RAG inference efficiency by maximizing KV-cache reuse through smarter caching, pipelined data transfer, and prefetching, significantly reducing latency (up to 2.47× faster TTFT).

Miscellanea

Thanks to Jon Barron for the website template.