Laika is a software-only onboard AI platform that runs inference directly on the flight computers already carried by small satellites, enabling in-orbit image filtering and classification without additional hardware. The system uses under one watt of power and can be deployed over-the-air to satellites already in orbit, helping Earth Observation operators avoid paying downlink costs for cloud-obscured or redundant imagery. Laika's proof of concept demonstrates cloud detection on a Cortex-M4-class processor in 0.93 seconds per scene, with a model small enough to fit in 128 KB of RAM.
- Aerospace
- Artificial Intelligence
- Software Only
- Space Technology
Funding
Founders
Product
Problem
Earth Observation operators pay full downlink costs for imagery they will ultimately discard—cloud-obscured scenes, redundant frames, and areas nobody requested. Data is typically filtered only after it reaches the ground, meaning transmission costs are already incurred for unusable content. With roughly 0% of Earth's surface covered by clouds at any moment, a significant portion of downlinked data provides no value.
Solution
Laika provides a software-only inference layer that runs on the microcontrollers already present in most small satellites, classifying scenes in orbit before transmission. The system filters, classifies, and prioritizes imagery directly on the flight computer, ensuring only data worth paying for comes down. Laika's runtime and models can be deployed over-the-air to satellites already flying, with no additional hardware, integration campaign, or added mass. The platform supports ARM Cortex-M and A-class processors down to 128 KB of RAM and under one watt of power, with models that can be swapped via OTA updates for different use cases like ship detection, wildfire monitoring, or change detection.
Target Audience
Primary customers are Earth Observation satellite operators and small-satellite constellation managers who need to reduce downlink costs and maximize the value of their existing space infrastructure.
Features
- Runs on existing flight computers (ARM Cortex-M and A class) with no accelerator board or added hardware
- Sub-watt power consumption (<1 W) versus 3–15 W for dedicated AI accelerator hardware
- Over-the-air deployment enables retrofitting satellites already in orbit
- INT8 quantized model is 9.2× smaller than FP32 baseline with only 0.1% accuracy loss
- Processes 144 patches per scene in 0.93 seconds on a Cortex-M4-class processor
- Model catalog supports multiple use cases including cloud detection, ship detection, wildfire monitoring, and change detection
- Bench-validated on customer's OBC image before delta-upload to satellites