The use of electromyography (EMG) for translating muscle activity into precise movements has become a pivotal technique to control myoelectric prosthetic devices, significantly improving the quality of life for amputees. Traditional EMG control of prosthetics relies on techniques that decode forearm muscle activity into discrete and simple gestures. Although robust and reliable, these methods are far from the dexterity of a real hand, and can hinder the users' acceptance. Recent advancements in neural network-based control offer a more intuitive approach, enabling amputees to achieve more natural control over multiple degrees of freedom in their prosthetics [1,2]. Similarly, Spiking Neural Networks (SNNs) used as a regressor, rather than a discrete classifier, can directly map EMG signals into kinematics, bringing additional advantages in terms of power consumption and latency. Combining this continuous real-time approach with mixed-signal neuromorphic circuits provides a light-weight and energy-efficient solution to be embedded in myoelectric prosthetics. We propose an SNN architecture compatible with neuromorphic hardware for encoding EMG signals into wrist kinematics. The training of this network is implemented with a continuous time spike-based delta-learning rule, co-designed with its corresponding neuromorphic circuits. Custom behavioral models were developed to simulate these circuits under realistic hardware constraints in NEST [3], allowing insightful evaluation of the learning rule and techniques increasing robustness. We trained the network (Fig. 1) with the HIT Simultaneous Control (SimCo) EMG dataset [4], measuring both EMG and kinematics for a variety of wrist gestures. The 8-channel EMG signal is band-pass filtered and converted into spikes through a layer of leaky integrate-and-fire (LIF) neurons, then connected to the basal dendrite of an output layer of two-compartment LIF neurons through plastic synapses. During training, the target kinematics of the associated movement are provided to the apical dendrite of each output layer neuron. The network develops a meaningful and robust representation of kinematics from EMG data. The software simulations account for the variability and fixed resolution of weights characteristic of the hardware implementation. On-chip learning mitigates performance losses from hardware constraints, exploiting the low-power advantages of analog neuromorphic circuits affected by device mismatch [5].