Abstract <p>Although neuromorphic computing achieves significant energy savings through spiking neural networks (SNNs), developing efficient methods suitable for hardware deployment for reinforcement learning in SNNs remains a challenging task. We present a new two-layer spiking neural network architecture for the actor part of the actor-critic framework. The main contribution is in the actor hidden layer that learns state representations via unsupervised spike-timing dependent plasticity (STDP), which clusters input states. The output layer is trained with temporal-difference-modulated STDP (TD-STDP) using the critic’s TD error. The actor hidden layer’s outputs can be directly reused by a spiking critic as fixed features, which allows cutting the number of critic parameters and enabling a fully local three-factor plasticity pipeline. Evaluated on the classic Acrobot and CartPole control tasks, our SNN-based actor demonstrates performance comparable to recent SNN baselines. Creating reinforcement learning methods based on local plasticity learning rules is important for future implementation on neuromorphic hardware.</p>