Space-Based Computing and Autonomous AI Are Rewriting Mission Science

On 29 September, IEEE Spectrum published a detailed look at how NASA and other space organisations are moving generative AI from Earth-bound laboratories onto spacecraft. The article, published at 13:23 UTC, describes a series of experiments that mark a shift in how mission science results are produced, validated and acted upon.

The newest piece of the dossier is a massive space-based computing constellation that STAR.VISION Aerospace unveiled on 25 September at the Global Digital Trade Expo. According to China in Space, the company plans to deploy 1,080 spacecraft by 2035: 720 dedicated to inference tasks and 360 to general compute. The project, called the Space.IDC1 Computing Constellation, aims to place significant compute capacity in orbit to take advantage of solar power and radiative cooling. STAR.VISION has secured 10 billion Yuan (1.49 billion US Dollars as of 28 September) in strategic credit from the Industrial and Commercial Bank of China and the Agricultural Bank of China, with at least 1 billion Yuan available for Phase One.

Autonomous systems leave the lab

IEEE Spectrum reports that NASA's Jet Propulsion Laboratory used Anthropic's Claude models last December to help plan two Mars drives for the Perseverance rover. In May, NASA and IBM placed a compressed AI model on the International Space Station and a satellite to identify floods and clouds from orbit. In July, astronauts on the ISS tested a large language model to see if it could help with maintenance procedures.

These experiments point to a larger shift in space engineering. For decades, engineers on Earth decided what a machine in space would do, and the machine did exactly that. Now researchers are testing whether nondeterministic systems like generative AI can give spacecraft more flexibility to interpret their surroundings, plan tasks, and one day make decisions for themselves.

Robert Ambrose, the former chief of NASA's Software, Robotics, and Simulation Division, told IEEE Spectrum that autonomy often does not have a deterministic outcome. How you got into a situation changes the behaviour. "So if you come into the same situation but from different paths, the outcome could be different. And so engineers hate that," he said. Ambrose spent much of his career working on autonomous systems at NASA, including autonomy for the Orion spacecraft and Robonaut 2, a humanoid robot that went to space in 2011. With Orion, he saw how quickly the testing problem could multiply. Engineers had to consider not just what the spacecraft might do, but all the different ways it could have arrived at a decision.

The hardware race on Earth and in orbit

STAR.VISION's constellation is not the first of its kind. China in Space notes that ADA Space and Zhejiang Lab's Three-Body Computing Constellation is ahead of all other firms, with a dozen satellites currently running in space. Orbital Chenguang has major financial backing, while Chaozhisuan Beijing Technology Co Ltd has well-linked partners. Both have demonstration satellites working in orbit too.

The company plans three spacecraft designs. The G1 verification craft will provide 40 to 60 PetaFLOPs of computing power by generating 11.5 kilowatts from its solar panels. The G2 standard model will host 160 to 310 PetaFLOPs through 48 kilowatts of power. The G3 flagship is expected to host 460 to 1000 PetaFLOPs by generating 140 kilowatts.

The company did not share details on orbits, inter-spacecraft links, radiator sizes, or mass. International regulatory paperwork that would provide further insight is yet to be filed as well. At the unveiling, STAR.VISION mentioned that it has booked an undisclosed number of launches aboard Orienspace's Gravity-1 and Galactic Energy's Ceres-2 to deploy G1 spacecraft. It did not suggest a possible target launch month or quarter. The company has signed agreements with the Hangzhou Institute for Advanced Study, UCAS, the Hangzhou Research Institute of Xidian University, Hongqing Technology, Baidu AI Cloud, and MinoSpace. Unspecified support has come from the governments of Deqing Moganshan High-Tech Zone, Fuyang Economic and Technological Development Zone, and Yunqi town, all in Zhejiang province, as well as Xichang in Sichuan province and Chongzhou in Guangxi.

Testing the limits of AI in the field

While space agencies push autonomy, independent researchers are testing how well current AI models perform on tasks that mission scientists might delegate. A blog post published on 29 September by Labqoat describes a benchmark called Wildlife CSI, which tested eight models on 2,000 photos of animal traces: droppings, footprints, feathers, eggs, and bones. The best model, Claude Opus 5.5, got the species right 37.15% of the time across the full set, according to the blog. Muse Spark 1.3 followed at 33.35%, then GPT-6 Astra at 31.70%. Claude Sonnet 5.5 came fourth with 25.15%. Even Opus got the family right on only 51.75% of photos.

Across all eight models, 842 photos went without a correct species identification. Opus identified 205 of 400 egg photos, but only 82 of 400 footprints. Across all eight models, 257 of the 400 footprints went without a single correct species guess. For eggs, that number was 123. The author notes a caveat: some species have more photos than others, so getting those species right adds more to the score. When scores are averaged by species so every species counts equally, Opus's egg score falls from 51.25% to about 37%. The author also notes that iNaturalist's own guidance says to identify only as precisely as the evidence allows, and that having a species name attached to an observation does not necessarily mean that species can be worked out from a single trace photo.

The model behind the curtain

On 29 September, a website called spacebunnyalpha.com published benchmarks for an anonymous reasoning model named space-bunny-alpha. The model has a one-million-token context window, accepts text, image and video input, and is listed on OpenRouter as of 24 September 2026. It is free during a preview period.

The site reports that on a standardized 60-question subset of GPQA Diamond, space-bunny-alpha scored 82.0%. For comparison, GPT-6 Sol scored 95.45% on an independent 198-question audit, GPT-5.5 scored 93.6% on OpenAI's published result, and MiniMax M3 scored 92.9% on an Artificial Analysis standardized evaluation. On MMLU-Pro, space-bunny-alpha scored 75%, while GPT-5.6 Sol scored 89.1% and Qwen3.8 Flash scored approximately 89.0%. On a 300-question subset of Humanity's Last Exam, the model scored 46.1% with a 95% confidence interval of 40.4 to 51.8%.

The site also reports that space-bunny-alpha used about one-third of Qwen3.8 Flash's output tokens on the same benchmarks: 305,989 output tokens versus 913,989. The model scored 10.0 out of 10 on data parsing and extraction and tool calling, but only 4.2 on general intelligence and 3.0 on trivia.

What it means for mission science

The dossier does not contain a direct link between the STAR.VISION constellation and the NASA autonomy experiments. But both point in the same direction: more compute in space, and more decisions made in space rather than on the ground.

IEEE Spectrum notes that the technology is still far from trustworthy enough to hand over control of a spacecraft. But engineers are starting to ask whether they can afford not to do so as missions become more complex, distant, and numerous. The article quotes Ambrose saying that engineers found ways to manage the complexity of autonomous systems, including automating the testing process. China in Space reports that the Central Government has designated the necessary technologies for space-based computing as a focus area for the 15th Five-Year Plan period (2026-2030), with the Ministry of Industry and Information Technology involved. The article notes that Chinese firms quietly obtained the lead in space-based computing amid the global artificial intelligence boom.

Meanwhile, the Wildlife CSI benchmark and the space-bunny-alpha results suggest that current AI models still struggle with tasks that require precise identification from limited evidence. The gap between what autonomous systems can do in a laboratory and what they can do in the field remains significant. For mission science, the implication is that AI will increasingly assist with data processing, scheduling, and anomaly detection, but human oversight remains essential. The experiments on the ISS and the Perseverance rover show that AI can help, but they also show that humans are still checking the work.