Nnodely - Neuralize Your Model
Gastone Pietro Rosati Papini, Alice Plebe, Mojtaba Sharifzadeh, Mattia Piazza, Sebastiano Taddei, Giovanni Scialla, Francesco Baroni, Davide De Martini, Giovanni Maria Francesco La Scala, Gioele Defrancesco, Filippo Faccini
AI summary
Problem
Purely data-driven neural networks struggle with interpretability and data efficiency for physical systems, while manually designing model-structured neural networks remains complex and inaccessible to domain experts.
Approach
The authors introduce nnodely, an open-source framework that provides modular, physics-informed building blocks and an automated workflow to seamlessly embed physical laws into neural architectures for end-to-end training and deployment.
Key results
- Outperforms classical regression in robotic friction estimation with an MSE of 0.054 rad²
- Achieves precise vehicle lateral tracking with a curvature MSE of 4.2 × 10⁻⁷
- Provides a modular library of interpretable components like FIR filters and fuzzy models
- Enables straightforward deployment via JSON, PyTorch, and ONNX exports
Why it matters
It lowers the barrier for robotics and control engineers to deploy accurate, data-efficient, and interpretable neural models on resource-constrained platforms.
Abstract
Modeling and control of physical systems remain challenging for purely data-driven methods, which often lack interpretability and fail to leverage prior knowledge. Model- structured neural networks (MSNNs) embed physical laws into neural architectures; however, their design and imple- mentation can be nontrivial. We present nnodely, an open- source framework that simplifies MSNN development through a modular workflow, improving interpretability, data efficiency, and deployment on resource-constrained platforms. The paper highlights the framework’s features, positions it within the landscape of existing tool, and demonstrates its effectiveness in two case studies. nnodely is released under the MIT license and is available at https://github.com/tonegas/nnodely.