Neuron Dynamics
Fully autonomous drones, robots, and satellites demand a new AI architecture – fast, low latency, low power, small footprint, highly performant. The Neuron Dynamics’ asynchronous, event-based AI processes only what changes in a scene — at the sensor, at milliwatts. Our low-SWaP AI delivers the low-latency perception, planning, and action needed to detect, track, identify and avoid threats, as well as execute complex missions.
PROGRAMS
Our mission
Neuron Dynamics builds asynchronous, event-based Physical AI for autonomous systems at the edge — drones, robots, and satellites. Our mission is to give these machines the fast, low-power perception, planning, and action they need to operate on their own in the real world. Our ATLAS platform, invented at the University of Michigan and licensed exclusively to Neuron Dynamics, processes only what changes in a scene — at the sensor, at milliwatts. We build for missions where power, latency, and bandwidth are the hard limits, from defense and space to commercial robotics.
Built for the edge
ATLAS — Asynchronous Technology for Low-energy AI Systems — was invented at the University of Michigan and is licensed exclusively to Neuron Dynamics. ATLAS pairs event-based sensors with State Space Models running on compute-in-memory hardware to deliver orders of magnitude higher performance at lowest energy and latency.
Event sensors are highly dynamic, low energy sensors that produce sparse streams of events versus traditional frames – the output of traditional RGB cameras and modern radars. For example, an event camera reports only intensity changes per pixel, with microsecond latency and 120 dB dynamic range that holds through direct glare and darkness. Computation happens only when an event arrives. There is no clock — energy consumption scales with activity in the scene, not with time.
Architecture adapted from Zhang, Hu, Lu et al., Nature Communications 17, 1513 (2026).
Three elements. One edge stack.
Event-based perception
A dynamic vision sensor fused with an RGB camera used sparingly for full-scene context. High dynamic range for detection into the sun or in darkness.
State Space Models
Real-valued SSMs that train with standard gradient methods, integrating event streams into temporally coherent tracks — vastly outperform contemporary transformer and related VLA/world model architectures at a fraction of the energy cost.
Compute-in-memory
Multiply-and-accumulate runs as analog matrix math inside a CIM (compute-in-memory) processor — removing the data-movement bottleneck that drives considerable GPU power draw.
SWaP is our North Star.
Contemporary AI was designed for data centers. For small, mobile autonomous systems it is too energy-intensive, too slow, and creates too much data. ATLAS addresses size, weight and power together.
| METRIC | TRADITIONAL CNN/DNN + RGB VIDEO | NEURON DYNAMICS ATLAS |
|---|---|---|
| AI processing power | 10 W (Jetson Nano) | ~10 mW |
| Latency, detection to classification | 100 ms+ | <1 ms |
| Data flow | 30–200 fps × 8 MP × 8 bits | ~95% less |
Space
On-orbit custody and change detection through eclipse and glare, reducing downlink and ground processing.
Drones
Fast-object detection and identification in poor lighting, without draining battery range or time on station.
Robots
Low-latency perception feeding motion planning for mobile and manipulating robots.
Neuron Dynamics in the news.
Research, awards and partnership announcements.
Zhang, Hu, … Lu · Nature Communications 17, 1513
The people behind ATLAS
Contact us
Please reach out with any questions — we’d like to hear from you.
Neuron Dynamics, Inc.
541 Jefferson Ave, Ste 100
Redwood City, CA 94063
Asynchronous Event-Based Physical AI for edge autonomy.
© 2026 Neuron Dynamics, Inc. All rights reserved.