Alvin Zhu | Projects

MIDAS Hand: Open-Source Tactile Dexterous Hand Platform

RoMeLa • (Apr 2026-)

I helped develop MIDAS Hand, a low-cost, anthropomorphic tactile dexterous hand for robot learning, data collection, manipulation, and sim-to-real research. The public release includes CAD exports, printable files, a parts list, assembly documentation, MuJoCo simulation models, and software repositories for control, retargeting, and teleoperation. The hand has 16 total DOF, 13 active DOF, 283 3-axis taxels, a weight of about 700 g, and a BOM of about $3k.

Website Code Video

Curriculum Reinforcement Learning with AI Agents and Parallelized Simulation

RoMeLa • (Oct 2024-Jun 2026)

I created a GPU-accelerated simulation and training pipeline for the humanoid robot BRUCE using MuJoCo-MJX, enabling massive parallel environments for curriculum reinforcement learning. Accurate dynamics modeling of complex mechanisms and backlash were crucial for sim-to-real transfer and policy robustness on hardware. AI agents, RAG, and statc schemas were integrated to iteratively construct and refine training curriculum. The policies were deployed onto our custom robot BRUCE, a kid-sized humanoid with 3 modes of parallel actuation.

Publication 1 Publication 2

Publication 1

Publication 2

Project SPLITTER: Foundational work for new Agile Modular Robot

RoMeLa • (Oct 2023-May 2025)

I created a GPU-accelerated simulation and training pipeline for the humanoid robot BRUCE using MuJoCo-MJX, enabling massive parallel environments for curriculum reinforcement learning. Accurate dynamics modeling of complex mechanisms and backlash were crucial for sim-to-real transfer and policy robustness on hardware. AI agents, RAG, and statc schemas were integrated to iteratively construct and refine training curriculum.

Publication 1 Publication 2 Publication 3

Publication 1

Publication 2

Robust Robotic Perception Systems for Humanoid Soccer

RoMeLa • (Jan 2024-Aug 2024)

I developed the full perception stack for the humanoid robot ARTEMIS, enabling full spatial awareness and advanced decision-making in dynamic RoboCup soccer environments. This stack integrates YOLOv8 with classical vision algorithms and point clouds for object detection, 3D pose estimation, and trajectory prediction robust to noise.

Page Publication 1 Publication 2

Object Segmentation using Vision Transformers and Deep Learning models

RoMeLa • (Oct 2023-Dec 2023)

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 positions are extracted from an Intel RealSense D435 camera's point cloud for simultaneous locomotion and grasping.

PDF

Cost Efficient 3D Printed Robot Dog

Evodyne Robotics • (June 2022-Sep 2022)

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.

Page