Deep Space Autopilot: AstroForge Hands Mission Control Over to Onboard Neural Networks

Navigating deep space to reach distant asteroids remains one of the hardest engineering challenges in modern aerospace. Traditional space agencies like NASA rely on massive ground teams and heavy redundancy to run deep space missions. During the OSIRIS-REx mission to rendezvous with an asteroid back in 2018, NASA assigned up to 100 human flight controllers across every single eight-hour shift to manage spacecraft operations.

Private startups building commercial asteroid mining platforms simply do not have access to those massive public budgets. California startup AstroForge, which raised $50 million in venture funding to mine asteroids, is taking a completely different technical approach. The company built an autonomous control system named Solo, a custom transformer model designed to run spacecraft flight routines without relying on constant commands from Earth.

AstroForge plans to fly its first autonomous system in 2027 on a launch vehicle operated by Stoke Space. That mission receives backup funding from NASA and plans to collect scientific data about the sun. Most satellite operators avoid using neural networks for direct orbital navigation, relying instead on traditional hand-coded control math due to reliability concerns. In fact, satellite operators tested the very first neural network for orbital orientation just last year.

AstroForge decided to build custom flight intelligence after facing critical communication failures on earlier test flights. In 2025, the company launched its Odin spacecraft into deep space, but lost contact with the vehicle. Earth lacks enough deep space ground antennas to maintain continuous links with probes traveling hundreds of thousands of miles away, and available communication windows remain extremely small.

When ground teams lost control of Odin, co-founder and chief executive Matthew Gialich realized his team faced a massive financial trade-off. The company could spend roughly $200 million building its own ground station network around the planet, or it could build an intelligent model to fix craft issues directly onboard.

Armand Awad, head of flight software at AstroForge, led the effort to combine classic control algorithms with custom transformer models. The team trained its core system using flight data from individual hardware subsystems alongside stream inputs from roughly 2,500 onboard sensors.

This onboard intelligence layer isolates hardware anomalies automatically. If the spacecraft loses track of its orientation, the system can cross-reference power fluctuations with sensor glitches to pinpoint broken components, power-cycling affected hardware without waiting for ground instructions.

AstroForge is testing the Solo software in shadow mode during its upcoming DeepSpace-2 mission, scheduled to launch on an Intuitive Machines moon flight. The software will process real sensor data in the background without controlling hardware, allowing engineers to verify performance before turning over full operational authority on future asteroid runs.