Back to challenges

Autonomous debris detection
in deep space

Develop and train an AI model capable of autonomously identifying and classifying space debris using raw satellite telemetry and low-light optical data.
Read more Contact us

Challenge Overview

The proliferation of space debris presents a critical risk to active satellites and future deep-space exploration. Current detection methods rely heavily on ground-based radar, which is limited by latency and atmospheric interference. This challenge seeks to push the boundaries of on-orbit processing by moving the detection pipeline directly to the satellite.

The Goal

Develop a lightweight, high-accuracy neural network capable of identifying objects smaller than 10cm using purely synthetic and low-light optical datasets provided by the MARTINA consortium.

What is being tested

Algorithmic efficiency (FLOPs), inference speed on low-power hardware, and false-positive reduction in high-noise starfield backgrounds.

Assignment & Scope

Participants are required to design and build a complete inference pipeline capable of processing satellite data efficiently and reliably. The solution should cover the full workflow, from data ingestion and preprocessing to model inference and output generation.

Evaluation will not be based solely on accuracy. Equally important is how well the solution performs under the constraints of orbital edge-computing environments, such as limited processing power, restricted memory, and bandwidth limitations. Participants are expected to demonstrate robustness, efficiency, and scalability, ensuring that their approach can operate effectively in real-world conditions.

Task & AI Expectations

  • Design a CNN or Transformer-based architecture optimized for edge devices.
  • Achieve an F1-score of at least 0.85 on the validation set.
  • Maintain an inference latency of under 50ms per frame on a standard commercial embedded GPU.
  • Implement robust handling for extreme lighting variations and sensor noise.

Data & Technology

Explore the technical foundation of the challenge. The provided data mimics the hardware constraints and sensor profiles of the upcoming Argos constellation.

Satellite Data Type: L0/L1B Multispectral

Preferred AI Methods: Yolo-variants, ViT, Spiking Neural Networks (experimental)

Processing Pipeline

Data Acquisition

Raw multispectral imagery is captured by simulated LEO satellites.

Pre-processing

Noise reduction and normalization using minimal compute overhead.

Inference

Your AI model analyzes the stream to detect and track anomalies.