Robot Learning Researcher
I work on robot learning, embodied AI, and dexterous manipulation. I completed my MEng in Computing (AI & Machine Learning) at Imperial College London with First-class Honours, where my thesis studied language-guided, closed-loop imitation learning for tabletop manipulation.
Currently a Research Assistant at Shanghai Jiao Tong University, I work on VLA benchmarking and dexterous arm-hand manipulation pipelines. I am broadly interested in scalable Vision-Language-Action training, embodied agents, and contact-rich manipulation.
My current work spans VLA evaluation, dexterous robot data pipelines, language-conditioned control, and agent-based robot systems.
Tracking model architectures, training data sources, and post-training trends across the VLA landscape to clarify where current systems scale well and where they still depend on human supervision.
Building teleoperation, data collection, and evaluation infrastructure for real-robot dexterous manipulation, with a focus on arm-hand coordination and contact-rich tasks.
Designing closed-loop language-guided control systems that decompose tasks, recover from failures, and connect natural-language commands to manipulation policies.
Exploring how LLM/VLM agents can reason over robot state, select procedures, and reduce human supervision in embodied systems.
At Shanghai Jiao Tong University, I am building the teleoperation, data collection, and evaluation stack for a LinkerHand O6 + xArm6 platform. This work grounds the real-robot side of my current interests, especially contact-rich data collection, dexterous action spaces, and deployment-time evaluation.
My MEng thesis built a closed-loop system where an LLM planner decomposes natural-language commands into a sub-task DAG and a behaviour-tree controller orchestrates perception, policy execution, and automatic recovery. The project achieved 100% success on pick-and-place and a 2.4× improvement over a code-as-policy baseline.
I built and maintain an open-source leaderboard tracking 136+ Vision-Language-Action models across 6 benchmarks including LIBERO, CALVIN, Meta-World, and RoboChallenge. This project helps me systematically compare architectures, training data sources, and post-training strategies across the current VLA landscape.
RoboClaw is an open-source framework for embodied AI assistance with natural-language robot control, procedure selection, and ROS2 execution across simulation and real robots. It gave me direct experience with agent orchestration and autonomous data-collection loops, which continue to shape how I think about agent-based robotics systems.
VLA benchmarking (Evo-SOTA) and dexterous-hand manipulation research. Building teleoperation, training, and evaluation pipelines for real-robot arm-hand coordination.
Developed agentic capabilities for Westlaw Deep Research Agent including multi-agent workflow orchestration. Benchmarked and fine-tuned small language models for guardrail classification.
First-class Honours. Thesis on multimodal imitation learning for tabletop manipulation. Coursework in deep learning, computer vision, reinforcement learning, NLP, and robotics.
Built LLM-based pipelines for automated knowledge-graph construction on large-scale GPU clusters. Acknowledged in a Findings of ACL 2025 paper.
Outside research, I spend most of my time climbing. I summited Kilimanjaro, played on Imperial College's Men's 1st Basketball team, and share my home with three cats.
