Problem
Research questions and friction points this paper is trying to address.
Modeling friction torque in robotic joints accurately
Generalizing data-driven friction models to new dynamics
Adapting friction models with minimal training data
Learning-based adaption of robotic friction models
🤖 AI Summary
To address low modeling accuracy, poor generalizability, and heavy reliance on labeled data in friction torque estimation for robotic collaboration, this paper proposes a physics-informed, meta-learning-based data-driven online adaptive modeling method. The approach integrates a differentiable physical layer—incorporating Coulomb and viscous friction priors—with LSTM-based temporal modeling, Bayesian uncertainty quantification, and lightweight online gradient updates. It enables rapid cross-dynamic-condition model adaptation from minimal new-task data (few-shot). Evaluated on six robotic arm joints, the method reduces average prediction error by 62% and achieves deployment latency under 5 ms, satisfying real-time closed-loop control requirements. Key contributions include: (1) a physics-guided meta-learning framework; (2) an end-to-end differentiable modeling architecture; and (3) an efficient, edge-deployable online adaptation mechanism.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Residual learning adapts friction models efficiently
Combines base and residual networks for accuracy
Generalizes well with minimal training data
🔎 Similar Papers
No similar papers found.
