CV

Visiting Fellow

Harvard University, Computational Robotics Group

Advised by Prof. Heng Yang.

• Designed MoTVLA, a mixture-of-transformers framework that unifies generalist vision-language reasoning with fast motion decomposition for robotic manipulation.
• Developed diffusion-policy methods to improve language steerability and inference speed in robot skill learning.
• Explored scene graph-based visuomotor policies that combine GNNs and diffusion learning for long-horizon skill composition under visual distractors.

Research Assistant

University of Michigan, HDR Lab

AI-driven robotic structure evaluation with Prof. Xiaonan (Sean) Huang.

• Developed an EvoGym data generation pipeline for evaluating robotic structures.
• Applied large language models to guide robotic structure optimization.
• Improved simulation workflows by combining structured data generation with AI-driven evaluation.

Research Assistant

University of Michigan, MMint Lab

Robot skill learning via large-scale video representation with Prof. Nima Fazeli.

• Built a robot task-learning system integrating multimodal vision-language models.
• Implemented instruction-based semantic segmentation to improve robot perception and execution.
• Investigated latent-space clustering and phase transitions in long-horizon visual representations.

Research Assistant

Huawei Technologies, Noah's Ark Lab

• Upgraded the SMARTS multi-agent RL simulation platform and resolved more than 20 maintenance and improvement issues.
• Helped organize autonomous driving competitions built on SMARTS.
• Improved scene encoding for the Argoverse dataset with Deep Graph Library methods and achieved a substantial trajectory prediction gain.
• Proposed diversity and admissibility metrics for trajectory prediction and used them to benchmark representative models.

M.S.

University of Michigan

Robotics

• Advised by Prof. Nima Fazeli and Prof. Xiaonan (Sean) Huang.

BASc

University of Toronto

Electrical Engineering

• Worked with Prof. Chan Carusone on reinforcement learning for analog circuit design.

MoTVLA: A Vision-Language-Action Model with Unified Fast-Slow Reasoning

arXiv

Vision-language-action modeling with unified fast and slow reasoning for robotic manipulation.

ViSA-Flow: Accelerating Robot Skill Learning via Large-Scale Video Semantic Action Flow

arXiv

Accelerates robot skill learning with large-scale semantic action flow from video.

CRITERIA: A New Benchmarking Paradigm to Evaluate Trajectory Prediction Approaches

ICRA

Introduces a new benchmarking framework for evaluating trajectory prediction methods.