
I am a second year Master of Science in Mechanical Engineering at Carnegie Mellon University. I am currently doing research on Human-Robot Interaction in the Robotics Institute with the Robotic Caregiving and Human Interaction (RCHI) lab, focusing on assistive robotics and communication of robotic intent.
I previously graduated from the California Institute of Technology (Caltech) with a Bachelor of Science in Mechanical Engineering. During my time at Caltech, I conducted Robotic Machine Learning research with the Autonomous Robotics and Controls Lab (ARCL) and Robotic Hardware research with the Advanced Mechanical Experimental Robotics (AMBER) Lab.
Contact
Email: julian.millan729@gmail.com
Linkedin: www.linkedin.com/in/julian-millan-hri
Github: https://github.com/julian-millan
Hackaday: https://hackaday.io/jmillan
Research
Curent Master’s Research (August 2025 - Present)
My research with the RCHI Lab has focused on utilizing VLMs for robotic use in assistive tasks. I have published work in agentic simulation and am currently examining robotic memory storage for VLM-based robotic manipulation.
Training a Biped to Walk with Behavior Cloning and Reinforcement Learning (June - September 2024)
I worked with Caltech Professor Soon-Jo Chung on researching how to use machine learning to control a bipedal robot’s walking behavior in a novel direction. The novel direction chosen was to use behavior cloning on a capture-point inverse kinematics system and inverse dynamics system.
The concepts primarily used in this project were types of Machine Learning (ML). I used Conservative Q-Learning (CQL), a form of Offline Reinforcement Learning (RL), and Behavior Cloning (BC) to attempt to create a bipedal walking algorithm for the robot. The simulation would run a little slow using the simple BC policy format, so I used a different policy saving format known as an “mlpfile” which both improved simulation speed and model performance.
Further work included improving post-simulation data visualization, fixing the broken model training pipeline, collecting robust training data, reviewing and implementing CQL, and tuning/improving model parameters. The end result was a 20% improvement in model accuracy. Differences in the performance can be observed with the slower previous version and the newer faster version I designed. For more detail, the final report can be viewed here. (PDF)
Designing an Optimized Robotic Ankle for a Bipedal Robot (June - September 2023)
I worked with Caltech Professor Aaron Ames in the Caltech AMBER lab on robotic hardware design. This was my first experience in research, and I had the task of designing an actuated ankle for a bipedal robot with a passive toe joint. I gained a lot of valuable experience in understanding robotic hardware and how to apply my mechanical knowledge in a robotics sense. I leveraged my skills in FEA and 3D CAD to iteratively design the ankle to be lightweight and compatible with rotary motors. I learned about the value of the speed and precision rotary motors have compared to the convenience of linear actuators.