Hi there 👋 I am an EECS undergraduate at UC Berkeley. I study AI agents, agent evaluation, and robot learning.
I work at the Berkeley RDI Lab (under Prof. Dawn Song), where I am a co-first author of Agents' Last Exam (ALE). I also work with Zhuo Xu and Prof. Masayoshi Tomizuka on robot-learning data, and with Prof. Avideh Zakhor on UAV perception. Outside research, I am building Equile — the dry lab for AI scientists.
My current research interests:
- Can we reliably evaluate long-horizon agents? How do we build task environments and automated evaluators whose verifiable rewards are strong enough to both measure frontier agents and train them (RLVR)?
- What data do robots actually need to learn from humans? Can heterogeneous egocentric human demonstrations become as useful to robot learning as web text is to language models?
- What does it take to run agents as real infrastructure — sandboxed, reproducible, schedulable — and how does that change how we study them?
News
Selected for the Inspire Fellowship by rednote — one of the first 6 fellows worldwide ($100K grant).new
New preprint: ScienceIDE — turning scientific codebases into agent-learnable environments (with AItonomy Foundation).
Building a large-scale robot-learning data benchmark with egocentric human manipulation data (with Zhuo Xu and Prof. Masayoshi Tomizuka).
Agents' Last Exam is on arXiv, with an open evaluation framework for long-horizon, real-world agent tasks.
Started Equile — the dry lab for AI scientists.
Our robot delivery system integrating VLA and agentic workflows was demonstrated at CES 2026.
Publications
2026
Agents' Last Exam
ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments
Research Projects
Robot-Learning Data Benchmark
With Zhuo Xu and Prof. Masayoshi Tomizuka: a large-scale benchmark for robot learning. I clean and curate heterogeneous egocentric manipulation and human-object interaction data (EgoVerse, EgoDex, HOI4D) so that diverse human demonstrations become reliable training and evaluation resources for embodied systems.
UAV-to-UAV Detection & Tracking
With Prof. Avideh Zakhor: zero-shot UAV-to-UAV detection and multi-object tracking under high-speed motion and strong ego-motion, integrating optical-flow-based tracking with detection baselines.
Equile
The dry lab for AI scientists: complete computing environments and clusters where research agents run real computational studies — trusted scientific codes, scheduled runs, and a record of how every result was made.
Experience
Berkeley RDI / BAIR
Research Assistant · Prof. Dawn Song · agent evaluation & RLVR
Robot-Learning Data Benchmark
Robotics research · Zhuo Xu and Prof. Masayoshi Tomizuka
UC Berkeley EECS URAP
Research Assistant · Prof. Avideh Zakhor · UAV perception
Equile
Building the dry lab for AI scientists — reproducible computing environments where research agents run real computational studies
Starbot
Robotics AI Engineering Intern · VLA and robot delivery