Alvin Zhu
I am an incoming PhD student at Berkeley AI Research, supported by the NSF Fellowship. My research has spanned robotics hardware, robot learning, simulation, and perception, shaping my broader interest in physical intelligence for real-world robot behavior. I am also working at NVIDIA GEAR with Mengda Xu, Yuke Zhu, and Jim Fan. I previously studied at the University of California, Los Angeles '23-'26, working with Yusuke Tanaka and Dennis Hong at RoMeLa. Late at night, you may find me vibe coding or playing poker with friends.
Updates
- Jun 2026: Started my internship at NVIDIA GEAR Lab
- Jun 2026: My paper DexEXO on a dexterous exoskeleton for operator-agnostic demonstration and learning got accepted to IROS 2026
- Apr 2026: I received the NSF Graduate Research Fellowship (NSF GRFP)
- Jan 2026: My paper AURA on cross-embodied RL with autonomously generated curricula and self-improvement got accepted to ICRA 2026
-
Previous updates
- Jul 2025: My paper on mechanical-intelligent curriculum RL for robots with parallel actuation got accepted to Humanoids 2025
- Jan 2025: Two of my papers got accepted to ICRA 2025: Cycloid QDD with Learning-based Torque Estimation and Multi-Modal End-Effector with Force Sensing using Gated Networks
- Nov 2024: My paper on variable inertial attitude control for quadrupeds jumping through space got accepted to AeroConf 2025
- Jul 2024: I became the 2024 RoboCup world champions with RoMeLa's humanoid robot soccer team!
Selected Projects
Full Projects PageRobust Robotic Perception System for Humanoid Soccer
I developed the full perception stack for the humanoid robot ARTEMIS, enabling full spatial awareness in dynamic soccer environments. Deep learning models combined with classical CV and pure depth map detection fail-safes allowed ARTEMIS to score 45 goals in 6 seated matches, overthrowing the reigning champions 6 goals to 1.
Object Segmentation using Vision Transformers and Deep Learning models
I integrated the Segment Anything Model (SAM) vision transformer with custom YOLOv8 detection weights to provide 95% accurate segmentation of slide handles and stairs. The segmented object's positions are extracted from the Intel RealSense D435 camera's point cloud for use in simultaneous locomotion and grasping.
Cost Efficient 3D Printed Robot Dog
The robot dog project involved designing and developing a fully functional quadruped robot. I 3D modeled and manufactured the upgraded big dog, ensuring a compact and efficient mechanical design optimized for strength and cost efficiency. I implemented a PID control system integrated with an IMU to enable real-time balance and stability.