A Science Strategy for the Human Exploration of Mars (2026)
Chapter: Appendix F: Implications of Artificial Intelligence for Human Mars Exploration
F
Implications of Artificial Intelligence for Human Mars Exploration
Technology readiness levels (TRLs) for machine learning (ML) and artificial intelligence (AI) have been defined to ensure a principled implementation process for robust, reliable, and responsible systems following an engineering systems framework (Lavin et al. 2022). One challenge is that ML systems operating in planetary environments may encounter new environments for which there was no prior experimental or even synthetic training data. Consequently, these systems may experience a switchback to an earlier effective TRL. However, only certain types of machine learning systems are likely to experience these switchbacks, and this is also a characteristic of adaptability, enabling a system to achieve greater robustness or flexibility (Saleh et al. 2003). Some expectations for human Mars exploration based on current trends in ML are briefly summarized.
Ongoing advancements in graphical processing unit computational efficiency are propelling petaflop (one quadrillion floating-point operations per second) capabilities into compact form factors. New application-specific integrated circuits will enable hardware-accelerated ML systems that implement large language models (LLMs) and other ML models. It is likely that radiation-hardened versions of these chips will become feasible in the future, given adequate investment. This appears particularly likely given growing commercial interest in space, including cislunar and deep-space environments.
LARGE LANGUAGE MODELS
LLMs are models constructed upon a type of neural network known as a transformer (Vaswani et al. 2017). They predict tokens based on a set of input tokens. LLMs have become ubiquitous since the release of OpenAI’s ChatGPT in November 2022 and are rapidly approaching or exceeding human performance levels on many metrics (Jones 2024). For an overview of LLM development and future prospects, see Aschenbrenner’s article titled “Situational Awareness: The Decade Ahead” (2024).
LLMs are anticipated to integrate all human-generated digital information and knowledge in the near future. Furthermore, LLMs are becoming computationally feasible on local processors. Future hardware advancements are likely to enable models trained on comprehensive human data to execute locally on suitable hardware.
AGENTIC ARTIFICIAL INTELLIGENCE
Agentic AI refers to ML systems that analyze, learn, decide, and act to achieve objectives. A simple example would be utilizing an AI travel agent to book a flight.
The potential implications of agentic artificial intelligence are extensive. For instance, astronauts could utilize AI agents to manage the execution of complex tasks, providing a virtual team of support personnel to augment human capabilities during exploration missions.
Potential applications of agentic AI include autonomous navigation, real-time data analysis, predictive maintenance, resource utilization and tracking, climate and weather prediction, enhanced communication, health monitoring, robotic assistance, scientific discovery, and mission planning.
NONHUMANOID ROBOTICS
Automated driving technologies and their space applications will enable capital-rich autonomous operations on the Mars surface, from rover navigation (Verma et al. 2023) to helicopter flight (Balaram et al. 2021).
Remotely operated robotics are already in use on Earth for hazardous applications such as mining (Shimaponda-Nawa and Nwaila 2024) and could facilitate human exploration in more perilous settings (subsurface Mars) than most have previously envisioned for human space missions.
HUMANOID ROBOTICS
Humanoid robotics are experiencing rapid progress, with a notable current limitation being training. However, this is being addressed through the use of reinforcement learning, where humans train humanoids. Once models are developed for a specific task, synthetic data can be generated and utilized for model enhancement.
In the context of human missions on Mars, tasks that lack Earth equivalents or are challenging to accurately simulate on Earth may necessitate the deployment of humanoid robots. Human astronauts could train these robots to perform these tasks on Mars, with models being developed and trained on Earth using synthetic data. Ultimately, these models could be uploaded to the robots for continued operation after human return. This approach would enable humanoid robots to continue performing a subset of human-compatible tasks on Mars as well as others not suited to humans.
