AstroForge Plans Self-Flying Spacecraft With No Earth Backup
AstroForge plans to fly its Autonomy-1 spacecraft in 2027 under the control of Solo, its in-house autonomous control software built around transformer models, the AI architecture also used in language models. Co-founder and CEO Matthew Gialich said the company does not currently plan to fly radios that can receive from Earth on that mission, according to TechCrunch.
Solo is an attempt to move flight decisions from the ground onto the spacecraft. The stack pairs traditional control code with models trained on test data for subsystems including power generation and navigation, plus an intelligence layer trained on data from about 2,500 spacecraft sensors. Gialich and head of flight software Armand Awad described it as a layered system, with deterministic controls, meaning fixed rules that respond the same way every time, underneath and learned models handling subsystem behavior and higher-level reasoning.
Before Autonomy-1 flies, Solo will get a flight test in shadow mode aboard DeepSpace-2. That spacecraft is set to launch alongside Intuitive Machines' third moon mission by the end of 2026. Shadow mode works like an observer pilot. Solo will watch live data and generate commands without actuating the vehicle, letting engineers compare autonomous choices against ground-commanded operations on real deep-space hardware.
DeepSpace-2 is described as a low-cost, autonomous interplanetary spacecraft built to fly often, according to AstroForge. It is intended to travel millions of kilometers to rendezvous with a near-Earth asteroid.
AstroForge, a California startup founded in 2022, is backed with $56 million in venture funding to develop technology to mine asteroids. It has launched two prototype spacecraft, and both suffered anomalies that prevented most mission objectives. In 2025, it launched its Odin spacecraft into deep space but had difficulty communicating with it. Odin flew as a secondary payload on Intuitive Machines' IM-2 launch, according to SpaceNews. AstroForge prepared its second attempt at reaching an asteroid roughly 15 months after its first spacecraft was lost, and for the deep-space mission was in the final stages of designing the mission, optimizing flight trajectory, and testing flight software, avionics and ground systems.
Deep-space links have low bandwidth, high latency and drop out easily. A vehicle millions of kilometers from Earth cannot wait for ground teams to diagnose a power fault or a navigation error. Onboard fault detection, planning and recovery are requirements for any sustained prospecting campaign.
The broader context here is that spacecraft autonomy usually advances in small steps, with autonomy added for specific tasks such as optical navigation or fault protection while the ground keeps command authority. In my view, removing Earth-receive capability on Autonomy-1 is understandable as an engineering forcing function, but it concentrates risk. If Solo misclassifies a sensor signature or enters a bad planning loop, there is no ground override.
Worth flagging for practitioners is that shadow mode on DeepSpace-2 is the right pattern, familiar from aviation and autonomous vehicle work, where a new stack runs passively against a proven system before it takes control. The useful question is coverage. Test-data models for power and navigation transfer only as well as ground testing matches flight conditions, and 2,500 sensors create complex interactions that are hard to replicate before flight. Cruise, rendezvous proximity operations and any mining-related proximity work will stress different parts of the stack.
Looking at what this enables if it works, the payoff is cadence. Low-cost interplanetary buses that can fly often and operate without continuous ground staffing would change the economics of asteroid characterization. Prospecting targets could be surveyed in parallel, failures absorbed without losing a flagship program, and lessons fed directly back into the transformer models.
