Fan Xu

I received my bachelor's degree in Computer Science from UC Berkeley. Now I am working as a research assistant at MSC Lab, UC Berkeley.

My research focuses on enabling robots to acquire generalizable manipulation skills, with a particular interest in dexterous manipulation. I am interested in how robots can learn through vision, touch, and physical interaction. More broadly, my goal is to develop robots that learn efficiently and adapt to new objects, tasks, and environments.

I am fortunate to work with Prof. Masayoshi Tomizuka and Prof. Costas Spanos. I am mentored by Wei Zhan, Ruihai Wu, and Shangding Gu.

I love robots 🤖. I want to build intelligent robots that improve productivity and make people’s lives easier.

I am seeking Ph.D. opportunities in robot learning for Fall 2027.
Fan Xu outdoors with mountains and trees in the background
Email LinkedIn
Recent highlights

Our team won first place in both tracks (Brick Assembly and Industrial Board Assembly) at the RoCo Challenge at IROS 2026!

01Research & publications

OmniPiano logo
Dexterous manipulationCo-first author · Project lead

OmniPiano: Diverse Dexterous Piano-Playing Challenges for Standard, Robust, Safe, and Multi-Agent RL

Fan Xu1,*,†, Fan Yang2,*, Yifan Xu3,*, Ruiyuan Zhang4,*, Yimou Wu5, Nuo Yi6, Muning Wen2, Guang Chen7, Costas Spanos1, Shangding Gu1,2

* Equal contribution   † Project Lead

A unified benchmark for dexterous piano playing with one to five Shadow Hands. OmniPiano brings standard, robust, safe, and multi-agent reinforcement learning into a shared task family, with 912 task settings and 36 baseline algorithms for evaluating dexterous control, robustness, safety, and cooperation.

arXivProject websiteCodeTutorial
Original bottle affordance point cloud from V2AP, colored by predicted affordance
Robot learningManuscript in preparation·Second author

V2AP: Learning Affordance Posteriors for Robotic Grasping from Human Video Priors

Yanghong Liu, Fan Xu, Zehao Lin, Jiakai Chen, Haoran Lu, Zhengtao Yao, Yubiao Zhang, Masayoshi Tomizuka, Wei Zhan, Ruihai Wu

Human contact does not always translate into a feasible robot grasp. V2AP turns egocentric and third-person human videos into robot-usable supervision by verifying prior-guided grasps through motion planning and physics simulation. The resulting robot posteriors support learning affordance predictors and grasp generators for new objects.

Human video priorsRobot verificationLearned affordances

Method & real-world examples
V2AP pipeline: human video input, hand and object reconstruction, human priors, physics-verified robot posteriors, affordance learning, and real-robot deployment
Method. Human-video contact priors are grounded in robot embodiment and physics to construct posterior affordance supervision.
Six real-world robot grasping examples paired with colored affordance point clouds for a mug, bottles, and containers
Real-world examples. Robot grasps paired with the predicted affordance maps used to guide them.
AI benchmarkingData contributor

Agents’ Last Exam

I contributed tasks to the benchmark and am credited in the author list.

Agents’ Last Exam tests whether AI agents can complete sustained, multi-step professional work and produce results that can be verified. Developed with industry experts, the benchmark spans more than 1,000 tasks across 13 industry groups.

arXiv

02Experience

Bosch

Software Engineer Internship

–

03Honors & awards

1st Place in Both Tracks at the 2nd RoCo Assembly Challenge IROS 2026

Brick Assembly Track and Industrial Board Assembly Track

Team award · UC Berkeley BAIR / MSC Lab

Lab announcementCompetition announcement
2026