
What happened
AstroForge built Solo, a transformer-based autonomous control stack made in-house, to fly on Autonomy-1 in 2027 on Stoke Space's first rocket, backed by NASA, which is expected to gather scientific data about the sun.
Why it matters
If Solo works, it could let a startup spacecraft resolve its own anomalies without the large ground-controller teams and big antennas that NASA-style missions rely on, an approach most spacecraft autonomy avoids because of neural-network reliability concerns.
What to watch
The test hinges on whether Solo can handle anomaly resolution in shadow mode aboard DeepSpace-2, slated to launch by the end of 2026 alongside Intuitive Machines' third moon mission, before Autonomy-1 flies.
WHO IT HITSThis matters most to startup space operators and their flight-software teams, who lack NASA-scale ground networks. If Solo performs, they may get a cheaper way to run deep-space missions; if it fails, the case for onboard neural-network control stays unproven.
Summaries like this, in your inbox every morning.
AstroForge's move toward onboard AI follows two prototype spacecraft that both suffered anomalies preventing them from achieving most of their mission objectives, including the 2025 Odin launch into deep space, which the company ultimately could not control. The company's own explanation points to a hardware constraint: there are a limited number of Earth antennas big enough to reach spacecraft hundreds of thousands of miles away, and the communication windows are small. CEO Matthew Gialich framed the choice as a trade between building a ground network at a cost of around $200 million for five dishes around the world, or trying to remove that need with a model.
The technical approach is deliberately bounded. Armand Awad, head of flight software, said the company leveraged advances in transformer models driven by the frontier labs. The resulting stack combines traditional control algorithms, models trained on test data for specific subsystems like power generation or navigation, and an overall intelligence layer trained on about 2,500 sensors in the spacecraft. Gialich described it as constrained autonomy at a very low sensor input, rather than general spacecraft autonomy.
What the company is attempting would be unusual: most spacecraft autonomy depends on traditional control algorithms because of concerns about neural-network reliability, and the first use of a neural network to control a satellite's positioning in orbit took place just last year. The plan is to prove Solo in "shadow mode" aboard DeepSpace-2 before letting it take charge on Autonomy-1, and Gialich has said he does not plan to fly radios that can receive from Earth on that mission, though he acknowledged the team may talk him into it by the time they fly. Whether this becomes a real shift for low-cost deep-space operations likely hinges on how Solo behaves when an anomaly actually occurs.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, LINE, or Slack.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
DeepSeek CEO Liang Wenfeng told investors the startup aims to rely on Huawei chips to train its models

Arm estimates agentic AI could raise data-center CPU demand from roughly 30 million cores to about 120 million…

Shopify is partnering with Meta so Muse, Meta's AI shopping agent, can complete purchases at Shopify merchants…

Simon Willison released llm-typesafe 0.1a0, an LLM plugin adding support for TypeSafe AI's Jev model, installe…
NVIDIA released Isaac ROS 5.0, adding agentic workflows, Isaac Skills for setup and manipulation, and support…

Anthropic launched Claude Opus 5.5, saying it matches Claude Fable 5.1 on most tasks while costing about 40 pe…
