Error Amplification Limits ANN-to-SNN Conversion in Continuous Control

Spiking Neural Networks (SNNs) can achieve competitive performance by converting already existing well-trained Artificial Neural Networks (ANNs), avoiding further costly training. This property is particularly attractive in Reinforcement Learning (RL), where training through environment interaction is expensive and potentially unsafe. However, existing conversion methods perform poorly in continuous control, where suitable baselines are largely absent. We identify error amplification as the key cause: small action approximation errors become temporally correlated across decision steps, inducing cumulative state distribution shift and severe performance degradation. To address this issue, we propose Cross-Step Residual Potential Initialization (CRPI), a lightweight gradient-free mechanism that carries over residual membrane potentials across decision steps to suppress temporally correlated errors. Experiments on continuous control benchmarks with both vector and visual observations demonstrate that CRPI can be integrated into existing conversion pipelines and substantially recovers lost performance. Our results highlight continuous control as a critical and challenging benchmark for ANN-to-SNN conversion, where small errors can be strongly amplified and impact performance. Code is available at https://github.com/xuzijie32/ANN2SNN-CRPI.

AI models that control robots require a lot of computing power. Brain-inspired AI, known as Spiking Neural Networks (SNNs), can drastically reduce this energy consumption, making them highly attractive for battery-powered drones and wearable devices. Instead of training these brain-like models from scratch (which requires robots to undergo expensive and potentially unsafe trial-and-error) researchers prefer to take standard, well-trained AI (artificial neural networks) and translate them into SNNs. However, this translation process often fails in "continuous control" tasks, such as smoothly steering a car or moving a robotic arm. In this paper, we discovered why: tiny translation errors at each movement step act like a snowball rolling down a hill. Because these small errors are linked over time across decision steps, they quickly amplify and accumulate, causing the robot to drift completely off course and fail. To solve this, we developed a lightweight software mechanism called Cross-Step Residual Potential Initialization (CRPI). CRPI acts as a micro-corrector: it takes the tiny "leftover" electrical signals (residual potentials) from a previous movement and uses them to correct the next movement, breaking the chain of accumulating errors. As a result, our converted brain-like AI can control robots smoothly and accurately, recovering lost performance while maintaining incredible energy efficiency.