Li Zhongxuan

I am currently pursuing a PhD in Robotics at the Department of Computer Science, The University of Hong Kong (HKU). In parallel, I co-founded Omni-X Robotics, a startup building teleoperation and data collection systems for robot learning, where I serve as CEO. Prior to this, I completed an integrated MEng (with Bachelor's degree incorporated) in Electrical and Electronic Engineering at Imperial College London.

In addition to my academic training, I have extensive industry experience. At HKU, I collaborate with TransGP through the InnoHK initiative. Before starting my doctoral studies, I worked at Huawei on optical network optimization and planning, and earlier completed an internship at Ocado Technology in the UK.

Email  /  Google Scholar  /  Github  /  LinkedIn

Omni-X Robotics builds robot teleoperation master arms (OmniLink) and a data platform (OmniPipe), providing a one-stop toolchain from hardware to software for robot skill data collection and model training.

I hold a broad interest in a variety of topics in robotics. In particular, I work on robotics task understanding/planning. My ultimate goal is to develop robots that can learn, reason, interact and evolve in real-world settings. Besides my profession, I have a great fondness about history and philosophy. Some papers are highlighted.

We propose a self-supervised framework for video-based sewn piece counting that learns task representations and dynamics directly from unsegmented demonstrations, without manual action annotation. The framework learns a task-state manifold via self-supervised visual pretraining, discovers discrete skills through information-bottleneck-guided vector quantization, and performs hierarchical probabilistic inference for robust progress estimation and verified counting on real-world factory data.

We present a failure-aware bimanual teleoperation framework that provides compliant haptic assistance while preserving continuous human authority. Learned purely from mixed successful and failed offline teleoperation data, a conservative success score modulates the assistance strength while a learned actor supplies the corrective direction, both rendered through a master-side joint-space impedance interface that steers the operator away from risky actions without overriding intent.

We survey world models for robotic manipulation, operationally defining a world model as an action-conditioned predictive system. The literature is organized by what is predicted (five representation families), how prediction connects to action (integrated prediction–action models vs. explicit predictive planners), and when prediction is used across pretraining, post-training, and inference.

We present UniBiDex, a unified teleoperation framework for robotic bimanual dexterous manipulation that supports both VR-based and leader–follower input modalities. The framework integrates heterogeneous input devices into a shared control stack with consistent kinematic treatment and safety guarantees, employing null-space control to optimize bimanual configurations for smooth, collision-free motion.

We propose a new topological latent control model for reactive human–robot collaborative manipulation of deformable linear objects. The method achieves real-time collaborative control and has been recognized as an ESI Highly Cited + Hot Paper.

Our paper was selected for Poster Spotlight at AIAAT 2025, highlighting our research on unified teleoperation framework for robotic bimanual dexterous manipulation.

As a key member of the Omni-Mani team led by Prof. Jia Pan (HKU), we secured the championship among 20 elite international teams.

Led by Prof. Jia Pan at HKU, our Omni-Mani team achieved first place in the deformable object manipulation challenge, demonstrating advanced robotics capabilities in cable management tasks.

This website design is inspired by Jon Barron