Michael Lutz

Hi! I am currently working with factor models as an intern at Citadel. Previously, I've conducted research at KScale Labs and Bardeen AI. Additionally, I am a student at UC Berkeley studying Electrical Engineering and Computer Science.

I am a member of Machine Learning @ Berkeley and am fortunate to have received the Z-Fellows grant. I am also passionate about mentoring high school students at BLAST AI, an AI education organization I helped found.

Email  /  Scholar  /  Twitter  /  Github

profile photo

Research

I'm fascinated by reinforcement learning, robotics, and memory hierarchy. Currently, I am most interested in researching end-to-end humanoid locomotion policies, highly-parallelized robotic simulation, and training a visually-grounded world model.

WILBUR: Adaptive In-Context Learning for Robust and Accurate Web Agents
Michael Lutz*, Arth Bohra*, Manvel Saroyan, Artem Harutyunyan, Giovanni Campagna,
arXiv, 2024
project page / arXiv

Training a demonstration ranking model to optimize context for black-box models in the web agent environment. Reached SOTA text-only results on the WebVoyager benchmark accross searching, travel planning, information extraction, etc.

Libraries & Projects

KSim: An MJX-based humaniod locomotion training library
project repo

A simple and efficient JAX-based library for highly-parallelized humanoid locomotion training in MJX environments. The library is designed to be easy to use and extend, and is optimized for training accross multiple GPUs and nodes.

Miscellanea

cs70 Academic Intern, CS70 Summer 2023

source code