StardustArm

StardustArm is a NASA-themed zero-gravity robotic arm simulator built with MuJoCo — a 6-DOF MEPhi arm on a spaceship reaching and grabbing floating target spheres, with both a hand-coded IK controller and a SAC+HER reinforcement learning pipeline. The hardest problems were getting the IK to actually converge (joint angles were only set as PD targets, never applied directly to qpos, so the arm would approach then drift away), hiding the obstacle spheres so they didn’t look like extra target balls, and keeping the ball from slipping and flying out of the single-jaw gripper. I’m most proud of the full-stack integration — damped least-squares IK, minimum-jerk trajectories, floating base thrusters, vision detection, obstacle avoidance, and RL training all working together for a cohesive on-orbit servicing demo. To test it, cd stardance, pip install mujoco gymnasium stable-baselines3 sb3-contrib tensorboard glfw PyOpenGL, then python main.py for the IK controller or python train_sac.py –run my_run –timesteps 500000 for RL training; left-drag orbits the camera, right-drag zooms, ESC exits.