Feedback ACT with Closed-Loop Dynamics for Trajectory Generation of Excavators Loading Tasks

Automating excavator loading tasks falls under the domain of robotic operations, often tackled using imitation learning. These methods rely on generating expert-like trajectories, followed by tracking through lower-level controllers. While motor-driven robotic arms have achieved highly sophisticated control, construction machinery like excavators struggle with the inherent complexities of hydraulic dynamics, which hinder accurate trajectory tracking. To address this, we propose Feedback ACT, a method that integrates a closed-loop dynamics network with feedback correction to generate trajectories that pre-compensate for controller inaccuracies. By supervising closed-loop outcomes during training and iteratively aligning predicted states, our method reduces end-effector imitation errors by 31% and decreases closed-loop failure rates from 84% to less than 12% in tests. This approach not only improves robustness against hydraulic nonlinearities but also ensures adaptability to distribution shifts, offering valuable insights for the practical development of autonomous excavation systems.