HAMS Simulation

What it is

My simulation work on HAMS, the Correll Lab's humanoid manipulation stack for the Unitree H1. HAMS bridges MuJoCo/RoboCasa and Isaac Sim to a ROS2 control stack over one shared CycloneDDS domain. It is a team project; my part was the simulation side.

What I did

• Compared four grasp-planning methods (centroid, PCA top-down, NVIDIA GraspGenX, and a ranked skill) head to head across three base conditions (frozen, hanging, standing free), scored with Wilson 95% confidence intervals.
• Found and fixed why a standing humanoid drops grasps: the old executor planned the reach in the pelvis frame, so the robot's sway dragged the target off the object. Re-anchoring execution to the world frame recovered success across the board (centroid went from 0/20 to 24/30 standing).
• Ran a posturography battery (mean sway velocity, medial-lateral RMS, sway-ellipse area, minimum margin-of-stability) to quantify how much each grasp method disturbs balance.
• Helped stand up a self-contained Apple-Silicon, CPU-only port so the whole humanoid sim runs on a laptop: MuJoCo rendering in software, ROS2 on FastDDS, and the viewers streamed to a browser over noVNC.

Why it matters

A humanoid sim that runs anywhere and behaves like the real robot lets us compare, debug, and measure grasps before they ever touch hardware. This work feeds a humanoid manipulation paper (publication pending, IEEE Humanoids 2026).

Part of HAMS, the Correll Lab's humanoid stack (a team project). Mentored by William Xie, PI Prof. Nikolaus Correll. Full case study: https://acwa-portfolio.netlify.app/#/projects/hams-simulation

Built With

• docker
• isaac-sim
• mujoco
• python
• robocasa
• robotics
• ros2