Hangxing Wei

I work on foundation models for embodied agents, focusing on latent action modeling, robotics, reinforcement learning, and vision-language-action models.

I received my M.Eng. and B.Eng. degrees in Cyber Science and Engineering from Wuhan University. I am currently a research intern at Microsoft Research Asia, working with Dr. Li Zhao.

News

2026

villa-X was accepted to ICLR 2026.

Research

My research focuses on learning action representations for embodied agents. I am particularly interested in how latent action spaces can bridge human videos, robot demonstrations, and vision-language-action policies, so that embodied models can learn from broader data sources and transfer more reliably to new tasks.

At a higher level, I aim to build embodied models that can acquire reusable action abstractions, reason over long-horizon behavior, and adapt across embodiments and environments with limited task-specific supervision.

I also work on AI infrastructure for robotics and maintain an interest in agent security, especially prompt injection and secure tool-use behavior.

Engineering

I enjoy turning repeated workflow friction into small, reliable tools. This includes research infrastructure such as azure_jobs, a lightweight Azure ML job submission CLI, and expr_tracker, a simple experiment tracking wrapper over local logs and online backends such as W&B.

I also like building personal utilities that make everyday computing more organized and pleasant. See Open Source for selected tools.

Selected Publication

ICLR 2026

villa-X: Enhancing Latent Action Modeling in Vision-Language-Action Models

International Conference on Learning Representations

Grounded latent actions in robot dynamics for VLA policy learning, improving simulation and real-world performance with zero-shot latent planning.

Paper · Code · Project Page