I'm a sencond-year PhD candidate in Computer Science at Beijing Institute for General Artificial Intelligence (BIGAI) and ShanghaiTech University. I'm fortunate to be advised by Prof. Ziyuan Jiao and Prof. Chenxi Xiao.

My current research centers on VLA and dexterous robotic hands equipped with tactile perception. I aim to enable robots to perform more complex and higher-level maneuvers.

In my previous research, I have primarily focused on human body motion, particularly interactive hand motion, which has greatly contributed to my current work. I feel incredibly fortunate to have had the guidance of Prof. Lan Xu, Prof. Jingyi Yu and Prof. Jingya Wang.

Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models

Qwen Team ( co-first author & core contributor)
Website  •   Arxiv  •   Blog

Alignment unlocks scale, and scale unlocks generalization. Qwen-RobotManip aligns heterogeneous robot embodiments into a unified Vision-Language-Action framework for scalable multi-source robot learning. It substantially outperforms prior models across all evaluated OOD settings and ranks 1st on the RoboChallenge Table30 v1 Generalist Track, demonstrating strong generalization across unseen tasks, scenes, and robot platforms.

Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System

Qwen Team
Website  •   Arxiv  •   Blog
Qwen-RobotNav is a scalable navigation foundation model built for agentic navigation systems, with a parameterized interface that dynamically controls task modes and observation strategies. Trained on 15.6M samples, it achieves new state-of-the-art results across major navigation benchmarks and demonstrates strong zero-shot generalization in diverse environments.

Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation

Qwen Team
Website  •   Arxiv  •   Blog
Qwen-RobotWorld is a language-conditioned video world model that predicts physically grounded future visual trajectories across robotic manipulation, autonomous driving, indoor navigation, and human-to-robot transfer. Built on a Double-Stream MMDiT and large-scale Embodied World Knowledge, it ranks 1st overall on EWMBench and DreamGen Bench, while demonstrating robust zero-shot generalization and multi-view consistency.

Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments

Qwen Team (core contributor)
Website  •   Arxiv
A unified embodied foundation model extending Qwen’s VL stack to action and trajectory generation via a DiT-based decoder, unifying manipulation, navigation, and trajectory prediction.

Research