Researchers led by Mahdi Nobar have improved how precisely a Franka Emika Panda robot arm follows a moving object on a conveyor, despite an imprecise digital twin, measurement errors and noise, according to a paper on arXiv under review for IEEE Robotics and Automation Letters.
They added a correction learned in the digital twin by reinforcement learning to a conventional controller, and used a Kalman filter that compensates for delays to estimate the target's position from camera measurements that arrive late or out of order. In simulation and hardware tests, the error fell below a limit the authors calculated for a class of model-based controllers.
At conveyor speeds of 27.9 to 41.8 mm/s, mean tracking error dropped from 1.1 to 1.4 mm with the conventional controller to 0.3 to 0.7 mm.



