Astrofurg puts its spacecraft under AI control with a Solo transformer model that eliminates ground-station costs

Deep-space explorations impose an operational and financial equation of great complexity; reaching a distant asteroid and working in its orbit historically required massive ground control rooms, as was the case in NASA’s 2018 “Oberon Rex” mission, which needed about a hundred human operators per shift. These huge resources remain a barrier for startups, and building a private ground communications network with five giant dishes worldwide costs roughly $200 million, not counting ongoing maintenance and operation expenses. In light of this, our platform followed Astrofurg’s decision, the asteroid-mining company, to forgo full reliance on ground control and develop a self-control stack powered by an internal transformer model it named “Solo,” to manage its spacecraft in space without direct guidance.

This shift did not arise from technological luxury; it was forced by a severe setback the company experienced after launching its “Odin” spacecraft into deep space in 2025. At that time the team faced harsh communication difficulties because of the scarcity of large ground antennas capable of transmitting over distances of hundreds of thousands of miles, coupled with narrow transmission windows, which ultimately led to loss of control of the craft, with the traditional onboard software unable to attempt autonomous fault correction. This failure pushed the company to leverage advances in transformer models, designing a stack that combines classic control algorithms and models trained on custom test data for subsystems such as navigation and power generation, topped by an intelligence layer trained on roughly 2,500 sensor readings inside the spacecraft to resolve emergent faults independently.

The experiment bets on demonstrating neural-network efficiency in open space, surpassing the sector’s historic caution that confined vehicle autonomy to rigid control algorithms for fear of generative-model errors.The practical testing of this pathway begins at the end of 2026 by sending the “Solo” model aboard the third spacecraft, “Deep Space 2,” accompanying Intuitive Machines’ third lunar mission, where the model will operate in monitoring and test mode without actual command authority. The next milestone will be in 2027 during NASA-backed “Autonomy 1,” a solar-study mission aboard the first rocket from Stoke Space, which the company plans to launch without any ground-based radio receivers, forcing the spacecraft to make survival decisions and manage its scientific payload relying solely on its computational mind.

We see in this shift a practical signal that redraws the planning of space projects and advanced systems in the Gulf and Egypt, where regional strategies have long relied on pouring massive capital investments into establishing ground control stations and tracking dishes that require continuous human operation. Demonstrating the viability of transformer-based control agents enables startups to slash deep-mission costs dramatically, shifting competition from owning heavy ground equipment to the efficiency of embedded software. And if you manage an engineering team or take part in developing local sensor systems, the most sought-after skill is no longer training large language models in data centers, but adapting edge models to handle thousands of sensors and physical data in harsh environments where communications are absent.

This trend sets a new benchmark for the autonomy of physical machines in general; a smart agent’s ability to infer faults and autonomously reset malfunctioning units meets a pressing need not only for spacecraft but also for robots and industrial-mining equipment operating in desert or remote areas cut off from digital coverage. The true test of autonomy software is not its theoretical performance in research labs, but its capacity to survive and manage crises in complete isolation without awaiting human intervention.