AstroForge Is Putting AI in Command of Its Next Spacecraft

Highlights

AstroForge is developing an onboard autonomous control stack called Solo, a transformer-based model intended to manage spacecraft operations without constant ground intervention. The startup plans a 2027 flight that will test Solo on a Stoke Space rocket, backed by NASA, and expects to collect solar science data. After earlier mission anomalies and communication challenges, the company opted to embed intelligence onboard to diagnose and repair problems in situ, rather than invest in an expensive global ground-antenna network.

Sentiment Analysis

• Overall sentiment: mixed but cautiously optimistic. The article balances excitement about novel AI-driven autonomy with sober recognition of past failures and technical risks. The tone acknowledges the potential benefits of onboard intelligence — reduced dependency on costly ground infrastructure and faster anomaly response — while also recognizing the maturity and reliability concerns that traditionally favor classical control systems. The narrative highlights AstroForge’s willingness to experiment and iterate after issues with prior prototype flights, suggesting resilience and learning. Given the cautious optimism, the sentiment intensity is moderate.

Article Text

AstroForge, a startup focused on asteroid mining technologies, is moving to place artificial intelligence at the center of spacecraft control. Faced with the high personnel and infrastructure demands of traditional mission operations — teams of flight controllers and a global antenna network — the company has developed an onboard autonomy stack, named Solo, built around transformer-style neural models. The intention is to enable spacecraft to handle tasks and resolve anomalies without relying on continuous ground intervention.

The company’s shift toward onboard intelligence was influenced by experience with earlier prototypes. Two small craft launched by AstroForge encountered problems that limited their missions; one recent deep-space vehicle, Odin, became difficult to contact. Deep-space communication depends on a small number of very large ground antennas and precise, limited-time communication windows. When those resources are scarce or a spacecraft becomes unresponsive, startups face hard choices: invest hundreds of millions in ground infrastructure or pursue smarter, more capable spacecraft.

AstroForge opted for the latter. Solo is a hybrid control stack that mixes conventional control algorithms, subsystem-specific models trained on test data (for power, navigation, etc.), and an overall transformer-based intelligence trained on thousands of sensor inputs. According to company leaders, the system is not meant to be a general-purpose AI for every spacecraft task, but rather a constrained autonomy tuned to the available sensors and mission profile. The goal is practical: give the vehicle the ability to correlate sensor readings, diagnose faults, and take corrective actions when needed.

One envisioned capability is anomaly resolution. For example, the onboard agent might detect a loss of attitude knowledge, trace that condition to a degraded star tracker or an associated power anomaly, and execute a recovery sequence such as power-cycling a unit or switching to redundant hardware. These kinds of decisions are typically handled by ground teams who analyze telemetry and send corrective commands; putting that capability onboard could shorten response time and improve survivability during long communication gaps.

To validate Solo in flight without fully relinquishing human oversight, AstroForge plans incremental testing. The company intends to fly Solo in a monitoring or "shadow" mode on an upcoming mission to exercise and observe its behavior under real conditions. Later, they aim to fly a mission in which Solo is the primary controller, with the ambition of operating without radios that receive commands from Earth during the flight. Leaders acknowledge that this represents a high-risk, high-reward approach: success would demonstrate that neural approaches can complement or replace some traditional autonomy techniques, while failure would underscore the challenges of relying on learned models in safety-critical space systems.

Technical and cultural hurdles remain. Spaceflight historically favors deterministic control algorithms because of their predictability and verifiability. Neural networks, especially transformer-based models, are relatively new to flight-critical control and raise questions about interpretability, verification, and robustness under out-of-distribution conditions. AstroForge’s architecture addresses some of these concerns by combining learned models with tried-and-true control logic, and by training on extensive sensor datasets. Still, the broader community will likely watch closely to see whether such systems can meet the reliability standards that deep-space missions demand.

The company’s fundraising, experience, and partnerships inform its strategy. Founded in 2022 and backed by venture capital, AstroForge has attracted attention for its ambition and for taking practical steps to iterate after setbacks. Collaboration with launch providers and support from agencies like NASA enable testing opportunities that might otherwise be unavailable to a small company. Those partnerships also provide external data points about how well onboard AI can perform when exposed to the complexities of real spaceflight.

Looking ahead, success would open a pathway for other small operators to pursue more independent missions without the prohibitive costs of global ground networks. It could also expand the range of missions that can rely on adaptive onboard intelligence to handle unforeseen situations. However, the path to that future will require rigorous testing, careful integration of traditional and learned components, and transparent reporting of both successes and failures. AstroForge’s experiments will therefore be an important case study in whether contemporary AI architectures can meet the demanding reliability needs of space operations.

Key Insights Table