Seeing the Unseen: AI-Powered Collision Avoidance with Event-Based Cameras

Space is crowded, and getting more so every day. From defunct satellites to rogue rocket parts, orbital debris is becoming one of the most urgent challenges for space sustainability.

At AIKO, we’re tackling this head-on with a powerful mix of AI and neuromorphic hardware. The goal? Smarter, faster, and energy-efficient onboard collision avoidance that doesn’t wait for Earth to intervene.
This work — developed in collaboration with the Università di Torino — was recently presented at the Space Operations Conference in Montréal, where it stood out among the best papers of the event. You can also dive into the technical details in our arXiv preprint.

Real-time Debris tracker | Credits LeoLab https://leolabs.space

A New Kind of Payload, Built for a New Kind of Problem

Our latest development is a collision avoidance payload that brings together two key ingredients:

• A bio-inspired, event-based camera
• A custom deep learning algorithm called Stack-CNN

The sensor doesn’t “see” like a traditional camera. Instead, it reacts to changes in light at the pixel level, making it ideal for spotting fast and untracked space debris. Combined with Stack-CNN, it enables real-time detection of threats and rapid execution of Collision Avoidance Maneuvers (CAMs), directly onboard.
No ground intervention. No delay. Just autonomy, where it matters most.

Why Collision Avoidance Matters

Since 1957, we’ve been launching more and more into orbit. The result? A growing cloud of debris — dead satellites, upper stages, and fragments from past collisions — that’s putting today’s missions at risk.
Here’s the breakdown:

• Large debris (>10 cm) can be tracked from the ground. If a potential conjunction is predicted, a Conjunction Data Message (CDM) is sent, and a CAM can be executed — if the satellite can maneuver.
• Small debris (<10 cm), however, often goes undetected from Earth. That’s where onboard systems become essential: detecting uncatalogued objects in real time, and reacting fast enough to avoid them.

Our system is designed to do exactly that.

What’s Inside: A Quick Technical Dive

The payload is built around two core technologies:
🔹 Event-Based Camera
A neuromorphic sensor that only captures changes in brightness, outputting streams of events with timestamps, locations, and polarities. That means:

• Temporal resolution up to 1 µs
• Dynamic range up to 140 dB
• Ultra-low latency and data rate
• Power consumption as low as 10–400 mW

It’s small, efficient, and built for the dynamics of space.
🔹 Stack-CNN Algorithm
This hybrid AI pipeline works in two steps:

1. Stacking: Frames are shifted and summed based on trial motion vectors. If the vector matches the debris velocity, the signal gets amplified.
2. Convolutional Neural Network (CNN) Detection: A convolutional network detects the object, estimates its speed, and flags any potential collision.

The result? An enhanced signal-to-noise ratio and a system that can spot even dim, fast-moving debris that others miss.

Under the Hood: How the System Works Onboard

Here’s what happens in orbit:

1. The event-based camera continuously listens to the environment, feeding a real-time stream of events to the onboard Processing Unit.
2. Stack-CNN kicks in, analyzing the data to identify any moving light signals across the sensor’s field of view.
3. When a trace is detected, a tracking algorithm (under development) estimates the trajectory and calculates the probability of collision (P).
4. If P exceeds a certain threshold, the system triggers a CAM via the propulsion unit.

All this happens autonomously, without waiting for Earth. It’s a closed-loop, low-latency safety system designed for the realities of modern spaceflight.

Real Data, Real Results

We tested our approach using a public dataset of sky observations from terrestrial event-based cameras. Among the recorded objects was a known rocket stage (SL-8 R/B), plus eight fainter objects.
The result? Nine out of nine were detected. Even better, our algorithm caught three traces that the original dataset authors missed.
That’s a strong sign that the Stack-CNN method is not only compatible with event-based data, but also more sensitive to subtle traces — exactly what we need in orbit.

What’s Next

We’re not stopping here. Upcoming developments include:

• Measuring detection sensitivity of the system based on debris size and field of view, to better understand reaction times for CAMs.
• Refining Stack-CNN to remove computational bottlenecks and boost onboard efficiency.
• Developing a fully neuromorphic pipeline — by converting the CNN into a Spiking-CNN and running it on neuromorphic hardware for ultra-low-power operation.

All of this is being developed in close collaboration with Università di Torino, blending academic research with industrial application.

Why It Matters

This technology is more than a research project: it’s a step toward a safer, more autonomous space environment. By giving satellites the ability to see and react to danger in real time, we’re protecting critical infrastructure and preserving orbital sustainability.
And beyond space? The fusion of AI and neuromorphic hardware we’re exploring here could ripple across domains — from robotics to autonomous vehicles to edge AI systems.
We’re building the tools to see the unseen. And we’re just getting started.

Get AIKO LAB’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Learn more: arXiv preprint | SpaceOps 2025 best papers