Framework
Within the Neu4mes project, the framework is the first main outcome: open-source software to design, train, validate, and deploy model-structured neural networks (MSNN) for modelling and control of autonomous mechanical systems.
The released implementation is named nnodely (“nn” can be read as “m”, forming Modely). It is the practical tool that supports the project methodology on real hardware and in the demonstrative case studies (wheel vehicles, quadruped robots, flying drones).
Purpose
The framework guides development from early experimentation through deployment on embedded platforms. In particular it provides:
- A library of neural models (NMs) and neural controllers (NCs) for components and complete devices, used as building blocks for full system models
- Tools to investigate different neural modelling and control strategies and to validate stability and performance of models and closed-loop systems
- Export of trained networks for embedded use (standalone C code, FPGA targets, and interchange via ONNX and Functional Mock-up Units FMU)
- Integration with external tools through APIs
The library is developed as an abstraction layer on top of existing deep-learning backends (such as Keras3.0 or PyTorch), so that the hybrid MSNN approach remains flexible and usable without rebuilding a general-purpose learning stack from scratch.
Development phases (PH1–PH6)
The nnodely development pipeline organises MSNN work in six phases from design to deployment, with iterative composition, retraining, and selective export:
| Description | Phase |
|---|---|
| Neural model definition — MSNN structure, inputs/outputs, and high-level specification (Neural Model abstraction) from mechanical principles and domain knowledge | PH1 |
| Dataset creation — experimental time-series (CSV, DataFrames), splitting, resampling, and automatic temporal windows | PH2 |
| Neural model training — managed training with custom losses, optimizers, recurrent networks, and closed-loop learning | PH3 |
| Neural model validation — analysis beyond standard ML metrics (e.g. FVU, AIC, residual and domain-specific checks) | PH4 |
| Model export — JSON, PyTorch modules, and ONNX for embedded deployment and third-party tools (selective export supported) | PH5 |
| Model composition — interconnecting modules via connect and closed-loop links; staged training and assembly of complex systems (including controllers) | PH6 |
Implementation
A preliminary version of the Neu4mes framework is already available under the name nnodely, distributed on GitHub under the MIT license (pip install nnodely). Tutorials, API reference, and case-study material are published in the online documentation.
During the project, the library is extended with models and controllers from the experimental domains and generalized for reuse by the scientific and industrial community.