Neuromorphic use cases
Eleven places event-driven, brain-inspired processing could fit in motion and sensing
Where we think neuromorphic and event-driven processing earns its place in physical hardware, tied to the motion concepts we have designed. These are ideas to test, not results.
What the good cases have in common
Each case below has at least two of three things: a sensor whose signal is quiet most of the time, a response that must come fast, and a power budget too small for an always-on processor. Each one is an idea for us to test. None has been built.
In our own motion designs
Hover monitoring on a levitating stage. A puck that hovers produces almost no new information. Event-driven estimation (our study N2) does work only when the puck moves or the field changes.
Slip detection in a stick-slip positioner. Each step is a short friction event. Spiking detectors suit catching a step that slipped badly before the position sensor shows it.
Part-present sensing in a zero-power gripper. The gripper uses no power to hold, so the electronics that watch it should idle too, waking only on a change in field.
Touchdown warning for a levitated pump. If the supply fails, the rotor has a fraction of a second before it touches. An always-on, low-power detector can tell a disturbance from normal running.
Assistive-device research with an electroadhesive clutch. Battery-powered research rigs that lock and release on intent signals need low-power, low-latency decisions.
Across instruments and machines
Event-camera alignment. Optical alignment watches for small motions. An event camera reports only changes, with microsecond latency.
Vibration watch on machine tools and pumps. Most of the time nothing is wrong. An always-on detector flags the change and wakes the main processor.
Battery-powered environmental sensors. A sensor that wakes on events instead of on a timer can last much longer on a battery.
Fast safety interlocks. Recognising a pattern in a sensor stream within a millisecond, close to the sensor, keeps the decision out of a slower main loop.
Adaptive control that keeps learning. Neuromorphic chips with on-chip plasticity can adjust a controller as friction or wear changes. In our stick-slip study, feeding back the recently measured step brought the biggest gain, which is a simple form of this.
Teaching. Spiking networks make timing, sparsity and energy visible in a way students can measure, on hardware that costs little.
Where it does not fit
Large, dense, always-busy models, such as today's language models, are better served by GPUs and dense accelerators. Intel's own lead researcher said in 2024 that they did not yet know how to map a language model onto Hala Point. The Loihi survey's finding that plain feedforward networks gain little points the same way.