Vision‑language‑action (VLA) models excel at robotic manipulation, yet their reliance on dense, continuous‑valued computation makes deployment on energy‑constrained hardware difficult. Inspired by the sparse, event‑driven signaling and temporal integration of biological neurons, researchers have explored spiking computation for VLA, but doing so without reshaping pretrained policies or breaking closed‑loop control remains challenging. To address this, we introduce the Power‑of‑Two Leaky Integrate‑and‑Fire (PoT‑LIF) neuron, which emits signed spikes whose amplitudes are powers of two. PoT‑LIF combines magnitude‑coded spike events with a residual‑preserving membrane dynamics, allowing continuous information to be progressively represented by discrete spikes.
The Heartbeat formulation further localizes this process within weighted projections, bridging biologically‑inspired spiking computation and existing pretrained VLA architectures without altering external inference interfaces. We embed PoT‑LIF into the vision projection, vision‑language projection, and action projection of (\pi{0.5}) and evaluate each configuration over 2,000 closed‑loop LIBERO episodes. In standard mode, PoT‑LIF VLA achieves success rates of 96.90% at (T=1) and 96.35% at (T=2); in Heartbeat mode it reaches 97.25% at (T=2) and 96.90% at (T=4), slightly surpassing the original continuous‑activation (\pi{0.5}) baseline of 96.70%. At (T=1) with a 60.01% firing rate, PoT‑LIF consumes roughly 1/7.67 of the dense‑reference arithmetic energy, corresponding to an 86.95% reduction.
Review: PoT‑LIF demonstrates that binary‑amplitude spikes can deliver efficient discrete information flow while preserving VLA performance, offering a practical spiking pathway for energy‑limited robotic systems.