Post by Filippo Moro

Our paper "Neuromorphic principles meet reinforcement learning for virtual robotic control" is out in Neuromorphic Computing and Engineering. I'm honored that it was selected for the journal's Rising Stars 2026 collection. Neuromorphic hardware promises low-latency, energy-efficient control, but training RL policies under neuromorphic constraints is painfully slow. We built a pipeline that makes it practical. We find: ⚡ Policy distillation (a strong teacher trained with SAC → constrained students + short RL finetuning) speeds up per-student training by >7.5× for ANNs and >8.1× for SNNs 🧑‍🎓 Distilled policies tolerate aggressive compression: 4-bit weights and up to 75% weight/activation sparsity with limited performance loss 🔀 DiffPOP, a differential population code for spike encoding/decoding, outperforms prior population-coding baselines while reducing MAC operations by 52% 🦾 Control is possible from event-based, delta-encoded inputs, but only with stateful policies. Sparsity, statefulness, and spike-based computation as three composable design axes: train one teacher, then iterate over students for whatever your target hardware demands. ❤️ Joint work with Melika Payvand (Institute of Neuroinformatics, UZH & ETH Zurich). Link in comments.