TL;DR
AstroForge has unveiled Solo, an in-house transformer-based AI, to run its 2027 Autonomy-1 spacecraft with zero ground control, a direct response to losing its Odin probe because Earth-based operators could not intervene in time.
What happened
- AstroForge announced Solo on September 22, 2026: a transformer-based autonomous control stack built entirely in-house for spacecraft operation.
- Solo layers three components: deterministic flight-control algorithms, subsystem models trained on power and navigation data, and a top-level intelligence layer trained on roughly 2,500 onboard sensors.
- Autonomy-1 launches in 2027 on Stoke Space's first rocket, backed by NASA, with CEO Matthew Gialich stating his current intent to fly with no radios capable of receiving commands from Earth.
- Solo will first fly in shadow mode on DeepSpace-2, launching alongside Intuitive Machines' third lunar mission before end of 2026, so engineers can observe its behavior before handing it full control.
- AstroForge, founded in 2022 with $56 million in venture funding, has lost two prior prototype spacecraft to anomalies, including Odin, which tumbled uncontrolled after a communications gap roughly 21 hours post-launch in February 2025.
Why it matters
- The economics forced the decision: building a proprietary five-dish global ground network costs roughly $200 million; replacing it with an onboard model costs a fraction of that, reframing AI autonomy as a startup survival strategy.
- Neural-network spacecraft control is nearly unprecedented: most spacecraft still rely on traditional deterministic algorithms, and the first neural-network-controlled satellite positioning happened only in 2025.
- No undo button exists: once Autonomy-1 separates from the rocket, any AI error and mission failure become the same irreversible event, making this one of the highest-stakes agentic AI deployments attempted.
- The Odin failure is the proof of concept for Solo: Odin was lost not to exotic hardware failure but to a communications gap that ground teams could not close fast enough, exactly the scenario Solo is designed to survive.
- Broader signal for enterprise AI: while most agentic AI deployments can absorb mistakes, AstroForge's architecture offers a stress-test model for what full autonomy looks like when rollback is physically impossible.
What to watch next
- DeepSpace-2 shadow-mode results before end of 2026 will be the first real data on Solo's anomaly-detection accuracy and the clearest signal of whether Autonomy-1 is viable.
- Whether Gialich holds the no-radio stance: his own admission that the team may talk him into adding receivers by launch day means the final autonomy level is still negotiable and worth tracking.
- Competitor and NASA responses: if Solo performs in shadow mode, expect other small-sat operators facing ground-network cost barriers to pursue similar onboard-AI architectures, potentially accelerating the shift away from traditional mission-control models.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.