Compliance Control Tuning in Dynamic Mechatronic Systems Assisted by Deep and Reinforcement Learning

Manipulation tasks often call for complex combinations of compliance control objectives such as trajectory tracking, energy efficiency, and smoothness, which cannot be jointly satisfied by a fixed-parameter impedance controller for different scenarios. Here we introduce a DL+RL solution to learn variable impedance control policies for a 7-DOF robot manipulator. We train an LSTM with multi-head self-attention module to refine reference trajectories with behavior cloning, and learn a PPO agent to continuously adjust per joint stiffness at runtime. Our physics-informed auto-damping formulation is based on critical damping theory which automatically links damping coefficients with stiffness, reducing the degrees-of-freedom of the action space while yielding mechanically principled impedance behaviour. We benchmark this method, trained on the DROID robotic manipulation dataset, against baselines comprising fixed-parameter impedance controllers, DL-only models and RL-only models. Our DL+RL method reduces control energy by 25.9% and motion jerk by 96% relative to the fixed-parameter baseline with statistical significance determined by paired t-tests. An ablation study highlights the benefit of each architectural choice, including our proposed physics-informed damping formulation and attention mechanism.