I am a master's student in the LAMDA Group at the School of Artificial Intelligence, Nanjing University, advised by Prof. Yang Yu. I started my master's studies in 2026, after completing my undergraduate studies at the same school (2022–2026).
My research interests lie in embodied intelligence, world models, and reinforcement learning. I am interested in how world models can support efficient policy learning and reliable decision-making for robots.
* Equal contribution. See Google Scholar for the full list.
arXiv preprint, 2026
U-GROW filters failure tails and prioritizes high-uncertainty states for world-model rollouts, improving the efficiency of VLA policy optimization without changing its training objective.
RSS Workshop on Diffusion for Robot Learning, 2026
UA-OPD uses teacher uncertainty to guide on-policy distillation, reducing reliance on unreliable teacher predictions when training embodied policies.
ICML, 2026
VLA-MBPO trains robot policies inside learned world models, combining consistent multi-view predictions with short, branched rollouts to make reinforcement learning more data-efficient.
ICML, 2026
A lightweight scheduler learns from offline data to accelerate existing robot policies, using a world model to evaluate faster action sequences while preserving task success.
arXiv preprint, 2026
An anticipation model adapts future subgoals as a task unfolds, guiding a goal-conditioned VLA policy through long-horizon robot tasks.
arXiv preprint, 2026
A perspective on evaluating world models through their usefulness for intervention, policy evaluation, planning, and policy improvement.
NeurIPS Workshop on Embodied World Models for Decision Making, 2025
A world-model framework for embodied decision-making that studies generalization across environments and the benefits of scaling training data.