Advancing Intelligent Personal Assistants for Human Spaceflight

For this study on IPAs for future human spaceflight missions to the ISS, Moon, and beyond, we selectively invited experts from a range of disciplines related to space operations and planetary exploration. Participants were recruited via purposive sampling based on their expertise in astronaut operations, training, or space technology development. Invitations were extended through professional networks, institutional affiliations, and referrals. The final sample included 15 domain experts, among them 3 active astronauts with a combined total of approximately 900 days in space and over 20 hours of extravehicular activity (EVA) on the ISS (see Table 1).

In addition to astronauts, the group included engineers with mission planning and crew training experience, instructors supporting astronaut training and EUROCOM operations, and scientists involved in lunar and Martian field studies. All participants reported prior use of conversational agents such as Alexa, Siri, Google Home, or ChatGPT, and several had direct experience with CIMON, the AI assistant deployed on the ISS.

Participants were invited to take part in semi-structured interviews conducted in a quiet room at our lab at the European Space Agency’s European Astronaut Centre (ESA-EAC). All participants provided informed consent prior to the interviews. The interviews lasted around 30 minutes. A pre-formulated interview guide was used to ensure consistency across sessions, while still allowing room for follow-up questions and deeper exploration of interesting topics raised by the participants. We observed that thematic saturation was reached by the 13th interview, with no new themes emerging in the final two sessions. The interview questions focused on the following areas:

• $\mathrm{■}$

  What are the essential requirements that IPAs must fulfill to be used in future space operations?

• $\mathrm{■}$

  What are the critical shortcomings of existing IPAs, such as Alexa, Siri, Google Home, and CIMON, that could hinder their effectiveness in space operations?

• $\mathrm{■}$

  Do the requirements for IPAs vary between missions to the ISS, the Moon, Mars, or extended-duration spaceflight? If so, how?

The interview guide was based on previous work in human-computer interaction, IPAs, and extreme environment operations (e.g., spaceflight, military). It was reviewed by two domain experts to ensure relevance and clarity. When participants raised unanticipated points, we asked follow-up questions to explore those aspects in more detail.

All interviews were conducted by a member of the research team familiar with the topic of space systems and human-agent interaction, which helped establish rapport and contextual understanding. Each session was audio-recorded and transcribed verbatim by the research team.

We conducted a thematic analysis following the six-phase approach described by Braun and Clarke (2006) [9]. This method was chosen for its flexibility and suitability in identifying patterns and meanings across qualitative datasets. The process included familiarization with the data, initial code generation, theme identification, theme review, theme definition and naming, and final report production.

Two researchers independently coded the transcripts using a manual approach (without software assistance). Code themes were initially developed inductively from the data, without relying on any pre-existing coding frame. After independent coding, the researchers met to compare their respective findings, resolve discrepancies through discussion, and consolidate them into a shared coding scheme. This collaborative process served as an informal approach to establishing inter-coder agreement, ensuring that themes were consistently and accurately represented.

Emergent themes were iteratively refined and synthesized into a coherent thematic structure that captured the key insights and challenges regarding the deployment of IPAs in human spaceflight contexts.

The interviews highlighted the importance of flexibility in IPA systems, emphasizing the need for adaptability across a wide range of tasks and contexts. This included not only practical applications such as checklist completion, suit monitoring, and navigation, but also the ability to modulate behavior based on different scenarios, emergencies versus routine tasks, and user preferences. Participants repeatedly pointed out that IPAs should offer a degree of personalization and responsiveness to unforeseen situations.

Scientist 4 envisioned a system that supports dynamic adaptation: “A personalized setup… that is also flexible depending on the activity.”

Instructor 2 framed adaptability in terms of learning from astronauts’ training performance: “If you know you will operate with someone, you will train with this person, spend some time with them, and adapt. At the moment, we’re only adapting the way we’re communicating with and treating astronauts. We don’t adapt procedures, for example. But procedures could be adapted using AI. During the training it might learn: ‘Hey, this crew member is always struggling with this one step.’ So when the crew member reaches this step, they could be shown a video of him or her performing it during training. Operationally, that could be very useful.”

A recurring theme was the need for a flexible communication style. Participants wanted to be able to tell the assistant how to behave. Engineer 4 noted that astronauts may prefer concise instructions: “Not verbose, or like very quick.”

Instructor 3 reflected on the possibility of training the AI to adapt emotionally: “Your Alexa you can also train to be more emotional or more rational. If you keep shouting at your Alexa and cursing at it, it also becomes very…[unfiendly]. But if you say ‘Hey, how are you this morning?’, then Alexa also starts responding in a more friendly, emotional, human kind of way.”

This personalization extended into entertainment and well-being. Engineer 5 described the importance of tailoring the assistant’s tone and content to the user’s mood and context: “It’s super personal. Some people really love it if they get jokes every five minutes. Others just want military-style instructions, without being bothered by any side-talk. Some might find it useful if their heart rate is up and somebody plays classical music. But others might think, ‘Stop this and let me work.’ That’s something the AI could do for you or offer you.”

While some participants were open to emotionally adaptive responses, many expressed discomfort with AI systems that mimic human behavior too closely. Scientist 1 and Engineer 4 raised concerns about systems attempting to simulate personalities or emotional intelligence without sufficient realism.

Scientist 1 shared: “I always find it super creepy when you try to make something more human-like. You know, these kinds of things where it looks too close to being real, but the technology is not good enough. I think because AI is the only thing that, no matter what, is trained on a subset of data. Whereas a person can take in all the information, and it’s also culture. It’s not flexible.”

Engineer 4 recounted his experience simulating AI characters: “My friend and I pretended to give our character to ChatGPT, but then you could always tell it was simulated. It wasn’t real. I would find that unnerving. Unless the AI is extremely good at fooling me or really consistent in its answers, I would rather not have a personality. It needs to be consistent, or it takes away the immersion.”

Engineer 5 echoed this sentiment, preferring machine-like behavior over human mimicry: “It would probably be a computer-style voice rather than pretending to be human. Right now, it is pretending to be human with feelings and emotions. If the AI knows you’re stressed, it will change its voice level like a human would. That feels wrong to me. It reads my heart rate and talks to me calmly. But if I’m relaxed and it knows I’m an hour behind schedule, it talks differently again. That might feel wrong, although maybe with time, I would get used to it.”

Although a few participants valued social and emotional cues in voice interaction, the majority favored a more neutral, logical assistant, especially in high-stakes or emergency contexts. Scientist 1 summarized this view: “If you’re actually in a life-threatening situation, you want precise answers. It’s okay not to be joking around too much.”

Overall, participants emphasized that flexibility should not only address task support but also adapt to user preferences, emotional context, and operational demands. However, this flexibility should be implemented with care to avoid overstepping into unrealistic or unsettling human imitation.

The importance of IPA’s reliability consistently surfaced as another key theme in the interview study. Drawing from their first-hand experience with AI deployments such as CIMON aboard the ISS and commercial systems like Alexa or ChatGPT, many participants expressed skepticism about the current technological maturity of AI assistants in spaceflight contexts.

Engineer 1: “Everything we’ve tried so far worked very poorly. Technologically, some of the solutions that we’ve used were of the classical type, where you had to train the system to recognize your voice. And even then, the performance was not great, because the system itself was really limited in computational power… It was consistently terrible. […] If you have a solution for operational use, it needs to work 100% of the time.”

This sentiment was echoed by another participant with engineering responsibilities for in-mission operations:

Engineer 2: “In space operations, things have to work. Not six times out of ten, not seven times out of ten – but ten times out of ten. If it works, then sure, but it needs to work reliably and consistently every single time.”

Several participants emphasized that current AI assistants often fail to meet those standards, citing issues such as long response times, unexpected behaviors, and inaccurate outputs. These shortcomings were viewed as particularly problematic in safety-critical settings such as EVAs.

Instructor 1: “The problem is, for the time being, if you’re completely relying on such an artificial intelligence assistant, you still have some unexpected behaviours. […] I’m not sure if I would put an AI assistant immediately in a [safety] critical environment, like an EVA or Moon exploration. I would rather put it in a support function, until it’s really reliable in the response.”

Scientist 1 added that the narrow training scope of most IPAs makes them ill-suited for complex, dynamic environments, suggesting that human crew members are still more capable of holistic judgment.

There was also a strong preference for human communication over AI in situations where ground contact remains viable, such as on the ISS. Participants highlighted the value of interpersonal trust, shared context, and linguistic nuance, all of which are difficult to replicate in human-AI interaction.

Instructor 2: “For an ISS EVA we have immediate direct communication. I’m sure the crew would always want to talk to their capcom, because they know them, they’ve trained together. […] Why would you not talk to someone on the ground and go to the AI? There would just be no advantage, it would only introduce a risk of miscommunication.”

Taken together, these insights suggest that while IPAs may play a supportive role in routine or non-critical tasks, current limitations in their reliability and predictability preclude their use as a substitute for trusted human interaction, especially in mission-critical scenarios.

The necessity for astronauts to operate independently from terrestrial mission control centers likewise emerged as a crucial requirement across interviews. Participants consistently emphasized that, in the absence of real-time communication, particularly during missions to the Moon’s far side or Mars, intelligent personal assistants must be fully functional offline.

Instructor 3 pointed out that the relevance of an IPA depends strongly on the mission context: “It largely depends on where your EVA is taking place. If it’s on ISS, then maybe not. But if you’re on Mars, you’d probably need something like this, because you won’t have real-time communication with the mission control centre.”

Engineer 1 raised technical concerns about embedding IPA systems directly into EVA suits, especially the challenge of implementing reliable AI locally without relying on Earth-based servers: “But how do you do that (integrating the AI assistant in the suit)? Because then you will need to implement the AI locally, it will need to be either in-suit, or it would have to be on a spacecraft. But ideally in suit, how do you implement something like that in suit?”

Beyond connectivity, the issue of hardware viability also surfaced. Engineer 1 noted the limitations imposed by energy consumption and computational power, particularly for radiation-hardened environments. Similarly, Astronaut 2 emphasized the importance of local control for mission-critical data and functionality:

Astronaut 2: “I want to have [local] control. I don’t want to have to ask the mission control centre, for example, to toggle information in my visor. Because I might not even have a connection to the mission control centre. So all this data, all this functionality, I want to have with me.”

In scenarios such as lunar far-side EVAs or deep cave exploration on Mars, crew members may need to retrieve localized fault diagnosis procedures or verify oxygen reserve levels without ground input. Throughout the interviews, multiple participants pointed out that in such cases, a failure of the IPA could risk the life of the crew. The assistant must thus support autonomous access to important procedural and diagnostic content.

Together, these perspectives underscore that offline operation is not merely a convenience but a mission-critical necessity. They point toward a need for self-contained, resilient, and locally executable IPA systems that can support astronauts autonomously under constrained conditions, without relying on cloud-based processing or external control infrastructure.

Most participants described voice interaction with an AI assistant as intuitive and hands-free, particularly well suited to the operational realities of space environments. Scientist 4, for example, drew inspiration from science fiction portrayals and explained: “I would prefer to talk… I wouldn’t like to have a lot of texts to read when you can just say it.”

However, several participants emphasized that relying solely on audio communication could be limiting, especially during tasks that require situational awareness or procedural clarity. The addition of visual information, either on displays or through AR, was described as important for maintaining efficiency and safety.

Astronaut 1 noted the risk of reducing information to the audio channel: “I think that I would not give up augmented reality. You know, voice can be an option to switch features on or off. But it might not always be safe [for more complex tasks] because you completely lose situational awareness if you are just waiting for a voice to respond to you.”

Engineer 4 highlighted the benefit of visual content for quick reference: “I’d rather be able to scan what it says.”

Engineer 3 emphasized the value of visual aids during procedural work: “You need procedures displayed somewhere… to show you the next three steps.”

Astronaut 3 envisioned an IPA that could integrate voice commands with augmented reality elements to support lunar surface operations: “Maybe you have Siri with you and you tell her to put up the map for today’s spacewalk, or the road to the next point. Then I would say, ‘Hey, can you put up in my head-up display a projection of the track?’ And later, ‘Okay, now it’s in the way, I don’t want to see it anymore. You can take it off.’ Things like that.”

These perspectives suggest that while audio-based interaction may serve as the primary modality for AI assistants in space, it should be complemented by visual output. This is particularly important for navigation, spatial orientation, and procedural tasks. AR displays were widely seen as a promising solution for presenting such information in a non-intrusive and context-aware manner.

Together, these four themes reflect a shift in how IPAs must be conceptualized within human-systems integration for space: not as convenience tools or as full replacements for human tasks, but as flexible, reliable, multimodal, and offline-capable systems designed to actively support astronaut-centered autonomy and decision-making during missions.

Overall, the study participants expressed a strong preference for IPAs that are highly adaptable in managing tasks whilst catering to a wide range of user needs during future human spaceflight missions. In contrast to terrestrial settings, where assistants support routine activities, future missions will demand on-the-fly adjustments to task sequences, unexpected procedural deviations, and adaptive communication styles.

For example, during an EVA, an astronaut may need to reprioritize steps in a checklist or receive contextual guidance for an unforeseen problem. Current rule-based IPAs are limited in these situations. Therefore, the participants suggested that there is a need for assistants that can understand intent, dynamically restructure procedures, and shift between brief, directive responses and more detailed explanations based on situational or personal needs.

Similar requirements have already been elaborated in other domains. For instance, in an effort to enable Unmanned Aerial Vehicles to better handle unpredictable mission scenarios, Keneni et al. advocated replacing traditional rule-based AI with a hybrid approach that integrates neural networks and fuzzy logic into a single system [20].

In the context of flexibility, prior work on human-systems integration (HSI) has mostly emphasized dynamic function allocation and adaptive control between human and automated systems [15, 35]. These approaches often focus on high-level decisions about when to delegate tasks or shift roles based on cognitive workload or mission state, giving the system a specific degree of flexibility depending on the situation. While valuable, this system-level framing partly differs from the type of flexibility highlighted in our study. As there is a difference between adaptable automation, where the human explicitly controls how automation is applied, and adaptive automation, where the system itself changes its behavior in response to inferred user or task states [11].

In our study, the participants described a need for interaction-level flexibility, namely, IPAs that actively adapts within tasks and requests to user preferences, communication styles, and situational needs, not only to the level of astronaut autonomy. Rather than passively taking over tasks, the IPA must respond to astronaut intent, restructure procedures dynamically, and provide multimodal feedback in a way that enhances and extends, rather than replaces, astronaut autonomy and active decision making, which would represent a form of adaptable automation, enhanced by adaptive elements that respond to inferred intent or context.

This aligns with findings in prior HSI studies, showing that adaptable automation leads to better task performance and reduced workload [11]. Given the safety-critical nature of spaceflight, astronauts may prefer the perceived control offered by adaptable systems, aligning with their desire to direct IPA behavior in context-sensitive ways (e.g., choosing terse vs. verbose feedback, or manually initiating a procedure switch).

Reliability was consistently suggested as one of the most important requirements for the use of IPAs in human spaceflight missions. Participants drew attention to past frustrations with unreliable assistants such as CIMON or Alexa, noting delays, misinterpretations, or incorrect outputs. In high-risk operational settings, such failures are not just inconvenient but potentially safety-critical. While users on Earth can fall back on human support when digital assistants fail, astronauts in deep space must rely solely on the system, making the error tolerance effectively zero. As such, participants emphasized that trust is a prerequisite for use. This aligns with prior research, which shows that trust in automation is closely tied to intended usage [12].

Participants also highlighted the importance of explainability, commonly defined as “the ability for the human user to understand the agent’s logic” [37, p.678]. Our research suggests, that being able to understand where a recommendation is coming from, how confident the system is, and what reasoning underlies its output could help astronauts to calibrate their trust in the IPA. This, in turn, helps avoid overreliance, while also ensuring that astronauts are not dismissing correct recommendations due to a lack of confidence in the system. These findings are supported by previous work on human-agent teaming and trust calibration [2, 37], and align with NASA’s ethical guidelines for AI use, which state that “Solutions must clearly state if, when, and how an AI system is involved, and AI logic and decisions must be explainable” [30, p.3].

Participants stressed that an IPA tailored for space missions would need to function offline, citing the need for local autonomy and reliability as key reasons. The connection between an AI system’s accuracy and its perceived trustworthiness is well established [25], and is further amplified in safety-critical contexts [46]. During operations on the Moon’s far side or on Mars, communication with Earth will be delayed or unavailable, resulting in a situation where IPAs will have to operate fully offline and deliver reliable functionality without any external humans support (e.g., from mission control centres).

Under such conditions, a system’s trustworthiness will inevitably be even more dependent on its perceived reliability. IPAs that fail or degrade under communication constraints could not be relied upon in critical operations. As such, the IPA must be resilient, self-contained, and capable of delivering consistent performance regardless of external connectivity.

The participants explained that relying only on audio information could reduce performance, for instance, when astronauts are hands-busy or have limited visual focus, additional output modes become critical. This observation aligns with Multiple Resource Theory [49], which suggests that human attentional capacity is limited and can be supported by distributing information across multiple sensory channels. Accordingly, multimodal interaction has been found to improve spatial orientation [10], safety [44], as well as learning performance of users [40]. AR technology integrated into the astronaut helmet could in this sense allow astronauts to quickly scan instructions, interpret navigation data, or check system diagnostics in a spatial, safe, and intuitive manner, reducing cognitive effort compared to relying on voice or text alone.

Participants described scenarios where audio feedback alone led to confusion or missed instructions. The combination of spoken dialogue with visual overlays could reduce cognitive load and could therefore support better situational awareness. Visually seeing the next three checklist items while hearing confirmations could also create redundancy and therefore reduce the chance of error.

Our findings, however, do not only show four identified requirements (flexibility, reliability, offline availability, and multimodal information) that should all be fulfilled for space-related IPAs, but also a tension between each of them. Fulfilling one could introduce additional friction for another requirement, requiring a balance between the identified factors.

For instance, participants expressed a strong preference for assistants that can dynamically respond to changing mission conditions or user needs (flexibility). Yet, making the assistant more flexible in its responses could create a higher risk of wrong or unexpected answers affecting its reliability.

Similarly, offline availability was framed as a non-negotiable requirement for missions beyond Low Earth Orbit. However, the need to operate without access to cloud-based resources raises questions about how to support and enhance explainability when the assistant cannot access external databases or justification mechanisms to allow for a high degree of flexibility.

We likewise came across instances of ambivalence surrounding the assistant’s interaction style. While the idea of a conversational agent that adapts emotionally to user stress or mood was welcomed in theory, many participants voiced discomfort with systems that attempt to mimic human behavior too closely. This was especially the case in situations involving risk or time pressure, where predictability and control took precedence over social or affective engagement.

Moreover, the integration of visual and auditory modalities was seen as critical to support astronaut performance in cognitively demanding environments. Yet this combination also introduces the risk of sensory overload if not carefully calibrated to task demands. The challenge lies not only in offering multiple channels of information but in ensuring that these are presented in ways that enhance rather than compete for user attention, while ensuring reliability of the system that in an ideal scenario could also react to as many novel situations as possible based on its database while also being available offline.

Therefore, all themes can be seen as important factors that should be in balance with each other to balance competing goals. A one-size-fits-all approach is unlikely to succeed. Instead, systems must be context-aware, adapting their behavior to the astronaut’s current needs and situation, workload, and environment while individually adjusting the specific requirement based on its importance in the given situation or mission.

To summarize, our findings indicate that future IPAs designed for space exploration must explicitly address the requirements of flexibility, reliability, offline availability, and multimodal interaction. The presented insights from astronauts and domain experts emphasize these important design criteria, exceeding the capabilities of traditional rule-based systems and narrow AI. Only the presented approaches (see related work) shown by METIS [6] and CORE [5], integrating KGs [7], generative AI (GPT models) [32], RAG [26], and, for instance AR, could be technical approaches to fulfill the requirements identified in our study (shown in Figure 2). Overall, our study reaffirms the importance of a human-centered approach to space technology development, what we define as a balanced astronaut-oriented design [4]. This approach prioritizes the astronaut’s perspective first, while iteratively involving all relevant stakeholders in a co-design process. Rather than retrofitting existing AI systems for extreme environments, IPAs should be co-designed with the end-user in mind, focusing not only on technical feasibility but on the real operational needs, constraints, and preferences of astronauts. Following this approach could also help with balancing trade-offs between the different requirement themes that emerged during the study, for example, because it enables iterative feedback from astronauts during the design process, developers can tailor the assistant’s interaction style (flexibility) to different task contexts without compromising trust (reliability), even in offline conditions where fallback options are limited.