Dexterous hand control requires low-latency adaptation to variations in object properties, contact conditions and interaction dynamics. Conventional reinforcement learning controllers are usually trained offline and require costly retraining when deployed performance degrades. This paper presents a reward-modulated neuromorphic accelerator that updates selected policy parameters on chip during deployment. A spiking neural network policy is initialized offline, while a reward-modulated rule enables online learning on chip by updating parameters from reward feedback and spike traces without backpropagation. A dataflow-aware dense-sparse architecture supports compute-dense encoding, event-driven sparse inference and local adaptation. Implemented in 40 nm CMOS, the accelerator occupies 3.74mm2, consumes 6.45mW at 100MHz, and achieves a 4.84kHz control rate. Under environmental changes, the grasping success rate with on-chip learning progressively increases over successive adaptation iterations, ultimately achieving an improvement of more than 10% over the fixed-policy baseline.
