Adaptive PID Control for Robotic Systems via Hierarchical Meta-Learning and Reinforcement Learning with Physics-Based Data Augmentation

Proportional-Integral-Derivative (PID) controllers remain the predominant choice in industrial robotics due to their simplicity and reliability. However, manual tuning of PID parameters for diverse robotic platforms is time-consuming and requires extensive domain expertise. This paper presents a novel hierarchical control framework that combines meta-learning for PID initialization and reinforcement learning (RL) for online adaptation. To address the sample efficiency challenge, a \textit{physics-based data augmentation} strategy is introduced that generates virtual robot configurations by systematically perturbing physical parameters, enabling effective meta-learning with limited real robot data. The proposed approach is evaluated on two heterogeneous platforms: a 9-DOF Franka Panda manipulator and a 12-DOF Laikago quadruped robot. Experimental results demonstrate that the proposed method achieves 16.6\% average improvement on Franka Panda (6.26° MAE), with exceptional gains in high-load joints (J2: 80.4\% improvement from 12.36° to 2.42°). Critically, this work discovers the \textit{optimization ceiling effect}: RL achieves dramatic improvements when meta-learning exhibits localized high-error joints, but provides no benefit (0.0\%) when baseline performance is uniformly strong, as observed in Laikago. The method demonstrates robust performance under disturbances (parameter uncertainty: +19.2\%, no disturbance: +16.6\%, average: +10.0\%) with only 10 minutes of training time. Multi-seed analysis across 100 random initializations confirms stable performance (4.81+/-1.64\% average). These results establish that RL effectiveness is highly dependent on meta-learning baseline quality and error distribution, providing important design guidance for hierarchical control systems.

Two-Step Image Dehazing with Intra-domain and Inter-domain Adaption

Recently, image dehazing task has achieved remarkable progress by convolutional neural network. However, those approaches mostly treat haze removal as a one-to-one problem and ignore the intra-domain gap. Therefore, haze distribution shift of the same scene images is not handled well. Also, dehazing models trained on the labeled synthetic datasets mostly suffer from performance degradation when tested on the unlabeled real datasets due to the inter-domain gap. Although some previous works apply translation network to bridge the synthetic domain and the real domain, the intra-domain gap still exists and affects the inter-domain adaption. In this work, we propose a novel Two-Step Dehazing Network (TSDN) to minimize the intra-domain gap and the inter-domain gap. First, we propose a multi-to-one dehazing network to eliminate the haze distribution shift of images within the synthetic domain. Then, we conduct an inter-domain adaption between the synthetic domain and the real domain based on the aligned synthetic features. Extensive experimental results demonstrate that our framework performs favorably against the state-of-the-art algorithms both on the synthetic datasets and the real datasets.