The problem
we solve
A satellite captures far more data than it can ever send down. Every downlink window is limited, and by the time raw data reaches the ground, the moment it was most useful for may already have passed. Transmitting everything wastes bandwidth on information that was never worth sending; waiting for the ground to process it wastes time a mission may not have.
Running AI directly on the satellite changes this equation, but it comes with a hard limit. Space-grade processors operate with a fraction of the memory, power and compute of a ground server, in an environment where nothing can be patched or replaced once in orbit. Intelligence that runs here is designed around that limit from the start, shaped by what the hardware can actually give it.
Use cases
_L0→L1C processing
Processing pipelines that turn raw sensor output into usable data products directly on the satellite, covering every stage from L0 up through L1B/L1C.
- CPU/GPU allocation: workload distributed across CPU cores, with suitable stages accelerated on GPU.
- Output readiness: data ready to use the moment it reaches the ground, or to feed a further onboard model.
_Semantic compression
Algorithms that determine, onboard, what is worth keeping and what isn’t before anything is transmitted.
- Mission-critical filtering: models trained to prioritise and preserve what matters, discarding the rest before transmission.
- Data volume reduction: downlink and inter-satellite volume reduced without losing analytical value.
- SoC-level optimisation: execution optimised for high-throughput on space-grade System-on-Chip architectures.
_Super-resolution
Algorithms that extract more detail from the same sensor, enhancing resolution through onboard processing alone. Together with semantic compression, super-resolution addresses the core trade-off of Earth Observation: reconstruction quality against transmitted data volume against onboard compute cost.
- Detail reconstruction: resolution enhanced from the same sensor input, with no new hardware required.
- Payload tasking support: sharper imagery feeds autonomous payload tasking and ground-scanning strategies.
- Real-time feedback: output can inform further onboard or satellite-level actions as it’s produced.
_Object detection
Models that recognise and track objects as data is captured, turning raw signal into an actionable result at the source.
- Onboard inference: detection runs without waiting for ground processing.
- Autonomous decision-making: the satellite can act on what it sees, or decide what to downlink, without a human in the loop.
How we do it
#Edge AI
#Computer Vision
#Adaptive AI
A shared architecture is the starting point for every deployment, then specified to the memory, power and compute budget of the target processor. In practice: a base design sized to the onboard accelerator’s footprint, followed by model inspection, quantization, compilation and inference through each hardware vendor’s toolchain. Where a platform offers different computing resources, reconstruction quality, processing performance and energy consumption are weighed against each other to decide how the model is allocated.
Experience spans Zynq-based platforms, Versal devices and Unibap hardware, validated on accelerators including Hailo-8/10, and AMD Radeon GPUs. Across all of them, the same discipline holds: footprint and latency reduced by orders of magnitude, accuracy kept where the mission needs it, proven on real hardware.
Hardware compatibility
Deployed.
Measured.
Proven.
ONBOARD SHIP DETECTION
95.7%
Faster inference on a Zynq-based space-grade board. Latency < 0.5 ms for a 2.5 × 2.5 km area.
ONBOARD SUPER-RESOLUTION
4x
Super-resolution with semantic compression on Unibap ix10. 40 → 10 cm/pixel.