CRITICAL safety skill for the WATonomous humanoid robot arm project. Use this skill ANY time code is being written, reviewed, debugged, or modified that will run on or could run on real robot hardware — including motor control code, CAN bus commands, ROS2 launch files, trajectory planners, teleop scripts, sim-to-real transfer, zero-position calibration, or any script that sends commands to CubeMars AK10-9, AK80-9, or GL40 motors. Also trigger when the user mentions "real arm", "real hardware", "deploy", "on the robot", "test on hardware", "run on the arm", "sim to real", "sim2real", "zero position", "homing", "calibration", "motor test", or any indication that code will command physical actuators. Even if the user just asks for a "quick script to move the arm", trigger this skill. AI-generated code for robot arms has caused physical damage before (stripped screws, broken mounts, violent jerking) — this skill exists to prevent that from happening again.
Use whenever the user wants autonomous or semi-autonomous iterative RL/policy training — launching training runs, watching metrics, diagnosing plateaus or divergence, tweaking reward/env/hyperparameter config, and repeating until a target behavior emerges. Applies to robot locomotion, manipulation, game agents, or any Isaac Lab / Gym / RSL-RL / rl_games style training loop where the user says things like "keep iterating," "train until it works," "watch this run and adjust," or asks for autonomous tuning of a training pipeline. Not just for humanoid walking — the loop generalizes to any sparse-reward, exploration-hard RL task.