CRIMSON-1: decide what to process in orbit

Singapore’s Space Access Programme, or SAP, supports opportunities to test technology in orbit within the wider Space Technology Development Programme. CRIMSON-1 launched on October 1, 2026, aboard Transporter-18 from Vandenberg. NTU’s next-day account identifies perovskite solar-cell experiments and onboard artificial-intelligence computing. The flight is a confirmed milestone; the announced experiments still require their own performance measurements and analysis.

The university’s account places the mission after earlier spacecraft including VELOX-AM. That history provides engineering experience without making the new experiment’s outcome automatic. Running a computation in orbit introduces power, heat and reliability constraints alongside the question of whether its output is useful. A smaller amount of returned data is valuable only when the retained information still supports the intended task.

Imagine a fictional instrument recording one hundred image tiles. An onboard model marks twenty as potentially useful and returns only those. The downlinked tile count falls by eighty percent, assuming equal tile sizes. But a ground review later shows that ten of the discarded tiles contained important features. The saving is real, yet the filtering decision has also removed evidence that the user needed.

This is a false-negative problem: the model rejected something that should have been retained. A system can also have false positives, sending irrelevant tiles and consuming bandwidth. Which error matters more depends on the application. A monitoring task might prefer extra irrelevant images to missing a rare important event. Thresholds should therefore be evaluated against the task’s costs and an appropriate, independently labeled comparison set.

Evaluate selection as well as savings

One test strategy is to return a sample of rejected data along with the selected data. That sample lets researchers estimate what the filter misses, provided the sampling method and uncertainty are understood. Returning only accepted examples would make the system difficult to evaluate: the ground team would see its apparent successes while losing direct access to many of its mistakes.

CRIMSON-1 also tests how computing workloads interact with the spacecraft’s heat and power conditions. A model that performs well on a laboratory computer may be unsuitable for a limited flight platform. The engineering question combines classification quality, processing cost and the communications benefit. SAP provides access to the relevant environment; the experiment must establish whether the complete tradeoff works under the conditions that matter to later users.