Data-Driven Dynamic Parameter Learning of manipulator robots

Data-Driven Dynamic Parameter Learning of manipulator robots Mohammed Elseiagy1,3, Tsige Tadesse Alemayoh1,2, Ranulfo Bezerra1,2, Shotaro Kojima1,2, Kazunori Ohno1,2 Abstract— Bridging the sim-to-real gap remains a funda- mental challenge in robotics, as accurate dynamic parameter estimation is essential for reliable model-based control, realistic simulation, and safe deployment of manipulators. Traditional analytical approaches often fall short when faced with complex robot structures and interactions. Data-driven methods offer a promising alternative, yet conventional neural networks such as recurrent models struggle to capture long-range dependencies critical for accurate estimation. In this study, we propose a Transformer-based approach for dynamic parameter estima- tion, supported by an automated pipeline that generates diverse robot models and enriched trajectory data using Jacobian- derived features. The dataset consists of 8,192 robots with varied inertial and frictional properties. Leveraging attention mechanisms, our model effectively captures both temporal and spatial dependencies. Experimental results highlight the influence of sequence length, sampling rate, and architecture, with the best configuration (sequence length 64, 64 Hz, four layers, 32 heads) achieving a validation R2 of 0.8633. Mass and inertia are estimated with near-perfect accuracy, Coulomb friction with moderate-to-high accuracy, while viscous friction and distal link center-of-mass remain more challenging. These results demonstrate that combining Transformers with auto- mated dataset generation and kinematic enrichment enables scalable, accurate dynamic parameter estimation, contributing to improved sim-to-real transfer in robotic systems I. INTRODUCTION Robotic arms are essential in today’s industries and re- search, carrying out tasks such as precision manufacturing, surgical support, and exploration of dangerous environments [1]. The performance and safety of these systems depend directly on the accuracy of their dynamic models. These models, defined by parameters such as mass, inertia, friction, and damping, determine how the robot moves and interacts with its surroundings. Estimating these dynamic parameters correctly is not just a theoretical exercise; it is a key require- ment for enabling advanced robotic functions and solving important real-world problems [2]. One of the main challenges in robotics is the so-called ”reality gap” in sim-to-real transfer. Training control poli- cies or reinforcement learning agents directly on physical robots is often difficult due to safety risks, limited time, and hardware costs. Simulations provide a safe and efficient alternative, but controllers trained in simulation often fail in the real world because of differences between simulated and real dynamics [3], [4], [5]. Accurate dynamic parameter 1 Graduate School of Information Sciences, Tohoku University, Sendai, Japan 2 Tough Cyberphysical AI Research Center, Tohoku University, Sendai, Japan 3 Egypt-Japan University of Science and Technology (E-JUST), New Borg El-Arab, Alexandria, Egypt mohammed.abelaziz@ejust.edu.eg estimation is crucial to close this gap. By carefully identify- ing a robot’s physical properties, we can build simulations that closely match real physics, making sim-to-real transfer more reliable and enabling advanced control and learning methods on real systems [6], [7], [8]. In addition, accurate models are the basis of digital twins, which are real- time virtual copies of physical systems used for monitoring, predictive maintenance, and optimization [9], [10], [11], [12]. For robotic arms, a high-quality digital twin allows safe testing of new control algorithms, early detection of wear and performance loss, and even remote fault diagnosis, improving efficiency and extending system life. Despite its importance, estimating these parameters re- mains challenging. Traditional analytical methods, such as least-squares estimation, often struggle with the complex and coupled dynamics of multi-joint arms [1]. These methods are sensitive to noise and can produce results that are not physically meaningful [2]. Moreover, dynamic parameters can change over time because of wear, temperature shifts, or changes in payload, requiring frequent re-calibration. Data collection in real robots adds further difficulty due to sensor noise and the challenge of obtaining accurate measurements. To overcome these issues, data-driven methods, especially those based on deep learning, have become strong alterna- tives. These approaches can learn system dynamics directly from sensor data without requiring explicit physical models. While Recurrent Neural Networks (RNNs) and Long Short- Term Memory (LSTM) networks have shown success in capturing time-based patterns [13], [14], [15], they often struggle with very long dependencies and may generalize poorly. The Transformer architecture, originally developed for natural language processing [16],

This content is AI-processed based on open access ArXiv data.