Haode Zhang 张皓德

I am a fourth-year undergraduate student at Shanghai Jiao Tong University, majoring in Mechanical Engineering. I'm currently working with Prof. Ruohan Gao and Ph.D. mentor Kelin Yu at UMD. I was also fortunate to work with Prof. Wanxin Jin from ASU and Prof. Yong-Lu Li from SJTU during my undergraduate.

Email  /  Scholar  /  Twitter  /  Github  /  LinkedIn

Research Interests

My research interests lie in robot learning, learning-based control and human-robot interaction. I want to build human-centered robots, robots that can:

(1) Learn from humans: leverage human effort and human knowledge efficiently — through explicit human feedback, implicit cues in human videos and the vast latent knowledge encoded in human language and datasets (VLMs and LLMs).
(2) Act like humans: incorporate bimanual tactile dexterous systems capable of performing much more complex tasks that could approach — or even surpass — human capability.
(3) Collaborate with humans: enable robots to collaborate with non-expert users, adapt to unstructured and dynamic environments, and personalize their behavior to diverse human preferences.

Feel free to contact me if we share same research interests.

I am currently seeking a Ph.D. opportunity for Fall 2026 in robotics!

News

[Jan. 2026] BSV is accepted to ICLR 2026.
[July. 2025] Start internship at UMD, College Park, Maryland.
[June. 2025] HSBC code is released. Check our page for more details.
[May. 2025] HSBC is accepted by ICML 2025.
[August. 2024] Start internship at ASU, Tempe, Arizona.

Publications

We propose a novel geometric view of reward alignment as an iterative cutting process over the hypothesis space. Our batched cutting method significantly improves data efficiency by maximizing the value of each human preference query. We introduce a conservative cutting algorithm that ensures robustness to unknown erroneous human preferences without explicitly identifying them.

We propose Scene Dynamic Field (SDF), a cost-efficient framework that integrates physics simulators into multi-task fine-tuning, substantially improving MLLMs’ intuitive physics understanding and achieving strong generalization across physical domains.

The website is built upon this template.