The reconstruction of motor neuron activity from surface EMG signals offers unique insights into human motor control and enables advanced human-machine interfaces that can be, e.g., used in prosthesis control. However, neural decoding is a complex problem, and despite decades of ongoing research, it is challenging to measure progress, as new algorithms are typically tested and validated on private datasets, using heterogeneous performance metrics.
To fill this gap, we introduce MUniverse, a simulation and benchmark suite for motor unit identification. This includes
- 6 diverse experimental and simulated datasets that are openly shared on Dataverse using the standardized BIDS format
- Curated and verified ground truth neural spike trains
- Easy-to-use EMG simulations for generating your own data
- 3 state-of-the-art motor unit identification algorithms that can be accessed through a unified API
- Simulation-based upper-bound predictions showing the unused potential of existing algorithms
- Standardized data formats and algorithm evaluation
MUniverse is presented at NeurIPS 2025, where you can find a short introduction video and the full paper.
The MUniverse project is a cooperation with researchers from the United Kingdom, Sweden, and France.
