An Interruptible Robot Arm

Most robot demos execute a single scripted command perfectly. Real operators change their minds halfway through, and almost no demo handles that. Change of Mind is a PyBullet-simulated Franka Panda arm driven by natural language. A command like "the red one needs to go to the bin on the left" is parsed by an LLM running on Groq into a structured pick-and-place task, then executed as six closed-loop cartesian waypoints. At any point during execution a new command can cut in. The arm stops, sets down whatever it is holding, a full gripper cannot start the next job, pushes the abandoned task's state onto a stack, and switches. Sending "resume" brings the original job back, re-finding the block wherever it was dropped rather than assuming it stayed put. Making that reliable took real instrumentation rather than guesswork. Measuring the simulator revealed that the gripper orientation used by the standard PyBullet Panda example inverts the hand on this URDF, driving the wrist 89 mm below the floor where it jammed, while inverse kinematics cheerfully reported a valid solution. Joint-limit arrays were being silently discarded because they covered 7 arm joints rather than the robot's full 9 degrees of freedom. The grasp constraint teleported blocks into the gripper link origin. Each was found by measurement and fixed. The result was verified end to end: all six object-destination combinations complete with under 2 mm placement error, zero missed grasps, and interrupt-and-resume tested at multiple trigger points including mid-carry.