AI Companions for Immersive Learning

From Conversational Bots to Adaptive Training Agents

A Case Study in AI, Virtual Reality, Education, Accessibility, and Experiential Learning

Prepared by MXTreality
Case Study: Europa Prime VR (EPVR)

Executive Summary

Artificial intelligence is rapidly changing how people interact with digital systems. Most educational applications of AI, however, still treat the technology primarily as a conversational interface: a student asks a question, and an AI provides an answer.

Our work with Europa Prime VR (EPVR) demonstrates a substantially broader model.

Within an immersive virtual environment, an AI agent can become an active participant in the learning experience; serving simultaneously as a companion, tutor, contextual knowledge resource, accessibility interface, gameplay guide, critical-thinking facilitator, and adaptive support system.

The AI companion developed for EPVR was designed to address a fundamental challenge in immersive education:

How can an AI help a learner without taking the learning away from the learner?

Rather than simply providing answers, the companion can provide carefully timed hints, contextual information, encouragement, questions, suggestions, and guidance. It can respond to what is happening in the environment and, when integrated with experience telemetry, potentially respond to how the learner is interacting with that environment.

This creates the foundation for a new class of educational technology: adaptive AI companions embedded directly inside experiential learning environments.

Europa Prime VR provides a practical demonstration of this approach. Developed through an NSF-funded research project in collaboration with TERC, EPVR was designed to broaden participation in informal STEM learning through virtual reality, with particular attention to neurodivergent learners.

The resulting architecture illustrates how AI can extend immersive learning beyond static scripted interactions into experiences that are:

• Conversational

• Context-aware

• Adaptive

• Supportive

• Knowledge-rich

• Accessibility-oriented

• Critical-thinking focused

• Data-informed

• Personalized

The implications extend well beyond entertainment. The same architecture can be adapted for workforce training, technical instruction, simulations, STEM education, healthcare training, onboarding, safety training, professional development, museums, and other experiential learning environments.

1. Introduction

Traditional training systems generally separate the learner from the instructor.

• A textbook provides information.

• A video demonstrates a procedure.

• A simulation provides an environment.

• A teacher provides guidance.

• An assessment measures performance.

• Artificial intelligence makes it possible to begin combining these functions into a single interactive system.

Immersive technologies such as VR add another dimension: the learner is no longer merely consuming information. They are acting within an environment. This creates an opportunity for AI to become something more sophisticated than a chatbot.

Instead of asking:

"How do we put ChatGPT into VR?"

the more important question becomes:

"How can an intelligent agent participate in an immersive learning environment in ways that improve learning, engagement, accessibility, and performance without replacing the learner's own thinking?"

Our work on Europa Prime VR explored this question in practice.

EPVR is a science-based VR experience set on Jupiter's moon Europa. The experience combines exploration, puzzles, scientific concepts, interactive systems, storytelling, and accessibility-oriented customization. It was developed as part of an NSF-funded project focused on informal STEM learning and broad participation, including neurodivergent learners.

Within that environment, the AI companion became more than a voice assistant.

It became part of the experience itself.

2. The Europa Prime VR Case Study

Europa Prime VR places the learner inside a fictional future research environment on Europa.

The experience uses real planetary science as its foundation while allowing speculative technologies and future scenarios to extend the experience into science fiction.

The project intentionally combines:

• Scientific knowledge

• Exploration

• Puzzle solving

• Interactive experimentation

• Storytelling

• Accessibility

• Adaptive settings

• Social interaction

• Discovery

EPVR's public description emphasizes that the experience is designed around real science while using science-fiction elements as extrapolations from current knowledge and evidence. This made EPVR a particularly useful environment for exploring AI.

An AI companion in such an environment has access to something a conventional educational chatbot does not:

context.

The learner is not simply asking an isolated question. The learner is:

• Standing somewhere

• Looking at something

• Attempting a task

• Solving a puzzle

• Making decisions

• Exploring

• Getting stuck

• Moving through an experience

• Potentially becoming frustrated

• Potentially losing attention

• Potentially succeeding

The AI therefore has the potential to understand the learner's interaction with knowledge as an experience, rather than as a sequence of questions.

3. From Chatbot to Learning Companion

The initial implementation demonstrated the ability to integrate a conversational AI directly into the VR experience.

The companion was given knowledge specific to Europa and the surrounding scientific context and was designed with a factual but approachable personality. It could discuss Europa, Jupiter, the Solar System, space travel, the surrounding environment, and other relevant topics while also interacting socially with the player.

This video demonstration shows several distinct functions:

1. Social interaction

2. Self-awareness of the environment

3. Guidance

4. Interactive suggestions

5. Additional knowledge

6. Integration with puzzles

7. Contextual greetings and interaction

This progression is important.

The AI did not need to be limited to:

"Ask me anything."

Instead, the agent could become part of the designed learning environment.

Video direct link https://youtu.be/SES8sQgbCZc?si=0ieXjR6xsC6doDAR

4. The Six Roles of an Immersive AI Companion

Our EPVR AI companion was designed to operate across six complementary roles, allowing a single AI agent to support the learner socially, educationally, and interactively while remaining part of the immersive environment.

4.1 Social & Emotional Companion

The AI can provide a sense of presence and companionship, helping prevent the learner from feeling alone in an immersive experience. It can greet the learner, engage in casual conversation, provide encouragement, acknowledge progress, and offer reassurance after unsuccessful attempts.

Example:
"That was a tough one. Don't worry—we can figure it out together."

4.2 Factual Science & Educational Knowledge Resource

The AI can provide accurate, contextual information about the subject matter being experienced. For EPVR, this included science and history related to Europa, Jupiter, space exploration, and other STEM topics. The system was designed around controlled knowledge and factual requirements rather than allowing the AI to freely invent answers.

Example:
"Europa's surface is primarily water ice, and scientists have strong evidence that a liquid ocean exists beneath that ice."

4.3 Critical-Thinking Tutor & Socratic Guide

Rather than simply giving learners answers, the AI can guide them toward discovering answers themselves. It can ask questions, provide graduated hints, challenge assumptions, and encourage learners to explain their reasoning.

Example:
"Before we change that setting, what evidence do you have that the receiver is the problem?"

This makes the AI a learning facilitator rather than an answer machine.

4.4 In-Experience Guide & Gameplay Assistant

The AI can understand the learner's current position and activity and provide assistance appropriate to the immediate situation. This can include puzzle hints, navigation guidance, explanations of unfamiliar objects, reminders about objectives, or suggestions about what to investigate next.

Example:
"You've already examined the power console. What other systems nearby might be connected to it?"

The important distinction is that assistance can be contextual and graduated, rather than simply revealing the solution.

4.5 Adaptive Observer & Engagement Support System

When connected to VR interaction telemetry, the AI can use observable behaviors—such as movement, interaction frequency, repeated attempts, time on task, and progression—to identify situations where additional support may be appropriate.

The system can potentially recognize patterns associated with:

• Increased difficulty

• Repeated unsuccessful attempts

• Loss of engagement

• Excessive inactivity

• Rapid progression

• Repeated requests for assistance

It can then adjust its behavior accordingly.

Example:
After detecting repeated unsuccessful attempts, the AI might say:

"You've tried that a few times. Would you like a small hint?"

This creates a potential feedback loop between learner behavior, AI intervention, and learning outcome.

4.6 Accessibility & Personal Comfort Assistant

The AI can also act as an interface to the experience's accessibility and comfort settings. Rather than requiring learners to leave the experience and navigate complex menus, they can communicate their needs naturally.

For example:

"The environment is a little too bright."

The AI could respond by recommending or applying available adjustments such as:

• Brightness

• Audio levels

• Environmental effects

• Motion settings

• Visual complexity

• Interaction preferences

This moves accessibility from a static configuration menu toward a more conversational and adaptive experience.

The Result

These six roles can operate simultaneously:

Companion → Knowledge Resource → Tutor → Guide → Adaptive Observer → Accessibility Assistant

The significance is that the learner does not need six separate systems. One appropriately designed AI agent can move between these roles based on the learner's context and needs.

5. Knowledge Without Turning Learning Into Answer Retrieval

One of the most important design challenges is factual reliability.

An educational AI cannot simply be treated as a general-purpose chatbot and assumed to be correct.

For science education, particularly in projects involving organizations such as NSF or NASA, factual accuracy and appropriate distinction between established knowledge, uncertainty, and speculation are fundamental considerations.

Our approach was therefore to constrain the AI's role around the knowledge appropriate to the experience.

The objective is not to claim that a generative AI model can never hallucinate.

Rather, the objective is to architect the system so that hallucination risk is reduced and unsupported claims are not presented as established fact.

This can include:

• Curated knowledge bases

• Domain-specific information

• Controlled system instructions

• Retrieval from approved sources

• Structured scientific references

• Separation of fact from speculation

• Explicit uncertainty

• Guardrails around unsupported claims

• Testing against known questions and failure cases

This distinction becomes especially important in a futuristic environment.

For example:

Established science

Europa has an ice shell and strong evidence supports the presence of a subsurface ocean.

Scientific inference

The ocean may have conditions that could be relevant to habitability.

Speculative future technology

Future human missions might develop technologies for sustained subsurface exploration.

These should not be presented as equivalent statements.

An educational AI can actually make this distinction part of the learning experience.

6. Teaching the Future Without Pretending to Know It

EPVR also explored an unusual educational opportunity: using current scientific understanding as a foundation for imagining possible futures.

A fictional environment can ask:

What might space exploration look like 50, 100, or 200 years from now?

The AI can participate in these discussions while maintaining a distinction between:

What we know

What scientists hypothesize

What is technologically plausible

What is speculative

What is purely fictional

This creates an opportunity to teach not only science, but scientific thinking.

Instead of saying:

"This technology will exist in 150 years."

the AI can say, in effect:

"Based on current developments in X, Y, and Z, this is one plausible pathway—but there are significant uncertainties."

This transforms futurism from prediction into an exercise in reasoning.

The learner can be encouraged to ask:

• What assumptions are we making?

• What technologies would be required?

• What scientific breakthroughs would be necessary?

• What would prevent this from happening?

• What evidence supports the idea?

• What alternative futures are possible?

The AI therefore becomes a tool for scenario analysis and systems thinking.

7. The AI as a Socratic Tutor

One of the most important principles demonstrated by the EPVR concept is that the AI should not always provide the answer. In educational games, giving the answer can undermine the purpose of the puzzle. A better model is graduated assistance.

Level 1 — Observation

The AI encourages the learner to notice something.

"Have you looked closely at the frequency display?"

Level 2 — Question

The AI asks a question that directs reasoning.

"What happens when you change the frequency?"

Level 3 — Hint

The AI provides a conceptual clue.

"Think about what the receiver is actually trying to detect."

Level 4 — Procedural Guidance

The AI explains the next type of action without completing the task.

"Try matching the signal on the left to the frequency shown on the monitor."

Level 5 — Explanation

Only when appropriate, the AI can explain the underlying concept.

This creates a scaffolded learning model.

The objective is not: 

AI solves the puzzle.

The objective is:

AI helps the learner solve the puzzle.

That distinction is fundamental to effective AI-assisted education.

8. AI-Assisted Critical Thinking

This model can be generalized beyond games. Imagine a training simulation in which a learner is troubleshooting a malfunctioning machine.

A conventional AI assistant might immediately say:

"Replace component X."

An educational AI could instead ask:

"What evidence makes you think component X is the problem?"

The learner identifies symptoms.

The AI asks another question.

The learner eliminates a possibility.

The AI introduces a new clue.

The learner eventually reaches the solution.

The system has therefore supported diagnostic reasoning, rather than merely delivering an answer.

This model can be applied to:

• Engineering

• Medicine

• Aviation

• Manufacturing

• Emergency response

• IT

• Science laboratories

• Automotive training

• Military simulations

• Safety procedures

• Technical maintenance

The core principle remains the same:

Support the reasoning process rather than replacing it.

9. Context-Aware Assistance

A conventional chatbot primarily understands what the user tells it.

An immersive AI agent can potentially understand what the learner is doing.

For example, the system may know:

• Which room the learner is in

• Which object they are examining

• Which puzzle is active

• Which controls they have used

• Which objectives they have completed

• How long they have been attempting a task

• Whether they have repeatedly attempted the same action

• Whether they have requested assistance

• What information has already been presented

This changes the nature of AI interaction.

The learner does not have to stop and ask:

"What am I supposed to do?"

The AI can recognize that the learner has reached a point where assistance might be useful.

For example:

"You've been working on that panel for a while. Want a hint?"

The system becomes contextually aware rather than purely conversational.

10. Behavioral Telemetry and Adaptive Learning

One of the most significant opportunities lies beyond conversation.

VR systems can capture interaction data that conventional classroom environments often cannot. Depending on the application's design and privacy framework, telemetry can include:

• Movement

• Interaction frequency

• Object selection

• Time on task

• Repeated attempts

• Progression

• Task completion

• Navigation patterns

• Pause behavior

• Requests for assistance

• Interaction sequences

This data can potentially be used to estimate states such as:

Engagement

Is the learner actively interacting with the environment?

Frustration

Are repeated unsuccessful attempts occurring?

Confusion

Is the learner repeatedly exploring the same area without progressing?

Attention

Is interaction declining or becoming highly fragmented?

Confidence

Is the learner progressing independently or repeatedly requesting assistance?

These measurements should be treated as behavioral indicators, not definitive psychological diagnoses.

That distinction is critical.

The system cannot necessarily know that a person is "frustrated."

It may observe behavior consistent with frustration.

This allows the AI to respond appropriately without pretending to read the learner's mind.

11. Adaptive Difficulty and Intervention

Once interaction telemetry and AI reasoning are combined, the experience can become adaptive.

Consider three learners approaching the same puzzle.

Learner A

Solves the puzzle quickly.

The AI remains largely silent.

Learner B

Struggles slightly.

The AI offers a contextual hint.

Learner C

Attempts the same action repeatedly.

The AI offers a stronger explanation or changes the presentation of the concept.

The underlying educational content remains the same.

The path to mastery changes.

This is one of the most powerful applications of AI in immersive education.

Instead of forcing every learner through an identical instructional sequence, the system can provide different levels and forms of support.

12. Accessibility as an AI Function

Accessibility is often treated as a collection of static settings.

1. Volume.

2. Brightness.

3. Subtitles.

4. Color.

5. Movement.

6. Controls.

These settings are important, but AI creates the possibility of making accessibility more dynamic.

EPVR was designed with extensive customization intended to accommodate different sensory, attention, and social needs. The project specifically emphasizes inclusive and customizable design.

An AI companion can become an interface to those controls.

Instead of forcing the learner to navigate a settings menu, the learner could say:

"The lights are too bright."

The system could respond:

"I can reduce brightness. Would you also like me to lower the environmental effects?"

Or:

"There's too much happening on screen."

The AI could recommend appropriate settings based on available options.

The result is an adaptive accessibility layer rather than a static accessibility menu.

13. Personalization Without Personal Judgment

Adaptive systems must be designed carefully.

The objective should not be to label learners.

Instead of:

"This player has poor attention."

the system can work with observable behaviors:

"The learner has paused interaction several times and has not progressed for four minutes."

That distinction protects the learner while still allowing the environment to adapt.

Potential adaptations include:

• Reducing visual complexity

• Adjusting audio

• Changing guidance frequency

• Offering additional hints

• Slowing the pace

• Repeating instructions

• Changing how information is presented

• Offering a break

• Increasing or decreasing challenge

This creates personalization based on interaction, rather than assumptions about the person.

14. The AI Companion as an Invisible Instructor

The most powerful version of this technology may actually be the least intrusive.

A traditional instructor often has to divide attention among multiple learners.

An AI companion can potentially provide each learner with individualized support simultaneously.

One learner might need:

"Give me another hint."

Another might need:

"Explain the physics."

Another might ask:

"Why does this work?"

Another might say:

"Just let me try it myself."

The same environment can support all four.

This does not eliminate the human instructor.

Instead, it allows the human instructor to move from answering every individual question toward:

• Designing learning experiences

• Monitoring progress

• Interpreting outcomes

• Providing higher-level intervention

• Supporting complex human needs

• Coaching

• Mentoring

• Evaluating mastery

AI becomes a force multiplier for educators.

15. Beyond Gaming: Training Applications

Although EPVR provides a game-based demonstration, the underlying architecture is applicable to professional training.

15.1 Industrial Training

An AI companion could guide a trainee through equipment operation.

It could:

• Explain components

• Ask diagnostic questions

• Provide procedural hints

• Monitor task progression

• Detect repeated errors

• Adjust assistance

• Record training outcomes

15.2 Healthcare Simulation

In simulated clinical environments, AI could act as:

• Patient

• Assistant

• Instructor

• Information resource

• Scenario director

Rather than simply telling a trainee what to do, it could respond dynamically to the trainee's decisions.

15.3 Safety Training

AI could support simulated emergency scenarios.

For example:

"The alarm has activated. What is your first priority?"

The learner makes a decision.

The simulation responds.

The AI asks the learner to justify the decision.

This creates active learning rather than passive compliance training.

15.4 STEM Education

Students could explore:

• Physics

• Biology

• Chemistry

• Astronomy

• Engineering

• Earth science

while interacting with an AI that can explain concepts in the context of what they are actually doing.

15.5 Workforce Onboarding

A virtual environment could replicate a workplace while an AI companion provides orientation, answers questions, and gradually reduces assistance as the employee becomes more proficient.

16. The Architecture of an AI Learning Agent

The EPVR experience suggests a conceptual architecture consisting of several layers.

Layer 1 — Immersive Environment

The VR/AR environment provides:

• Objects

• Spaces

• Tasks

• Characters

• Simulations

• Puzzles

• Scenarios

Layer 2 — Interaction Layer

The system observes relevant interaction events:

• Movement

• Object interaction

• Task progression

• Errors

• Attempts

• Requests for assistance

Layer 3 — Knowledge Layer

The AI accesses controlled information sources appropriate to the experience.

This can include:

• Scientific information

• Training manuals

• Procedures

• Curriculum

• Institutional information

• Approved reference materials

Layer 4 — Reasoning and Dialogue Layer

The AI interprets the learner's request and context.

Layer 5 — Pedagogical Layer

This layer determines how the AI should help.

Should it:

• Ask a question?

• Give a hint?

• Provide an explanation?

• Offer encouragement?

• Remain silent?

• Escalate assistance?

Layer 6 — Adaptation Layer

The system can modify the experience based on learner interaction.

Layer 7 — Analytics

Aggregated data can help educators and designers understand:

• Where learners struggle

• Which concepts cause confusion

• Which interventions work

• Where engagement declines

• How long tasks take

• Which learning paths are most effective

This architecture is significantly more powerful than simply embedding an LLM into a virtual character.

17. The Importance of Guardrails

AI in education must be designed differently from general-purpose consumer AI.

Several safeguards are particularly important.

Factual Guardrails

The system should distinguish verified information from speculation.

Pedagogical Guardrails

The AI should know when not to provide an answer.

Behavioral Guardrails

Telemetry should be interpreted cautiously and should not be presented as psychological certainty.

Privacy Guardrails

Learner data should be minimized, protected, and collected only for clearly defined purposes.

Transparency

Learners should understand that they are interacting with an AI system.

Human Oversight

Educators and administrators should retain meaningful control over learning objectives, content, evaluation, and intervention.

Failure Handling

The AI should be able to say:

"I don't know."

That may be one of the most important features of an educational AI.

A system that acknowledges uncertainty can teach learners that uncertainty is a normal part of scientific and professional reasoning.

18. Why Immersive AI Is Different

A conventional educational chatbot exists outside the learning environment. An immersive AI companion exists inside it. That difference matters.

A chatbot might explain how a hydrothermal vent works. An immersive AI can explain it while the learner is standing next to a virtual hydrothermal vent.

A chatbot can describe a machine. An immersive AI can explain the component the trainee is currently holding.

A chatbot can explain a procedure. An immersive AI can observe the learner attempting the procedure and intervene at the appropriate moment.

This produces a fundamental shift:

From information delivery to situated learning.

Knowledge is presented when and where it becomes relevant.

19. Situated Learning and Memory

Immersive environments offer the possibility of connecting information to action.

Instead of learning:

"Europa may have a subsurface ocean."

the learner can travel beneath the ice.

Instead of reading about a scientific instrument, the learner can operate one. 

Instead of memorizing a procedure, the learner can perform it.

The AI companion can then connect the experience to the underlying concept.

This combination of:

environment + action + feedback + explanation

can create a richer learning loop than information delivery alone.

The EPVR project was explicitly designed to investigate the affordances of VR for informal STEM learning, including engagement with diverse learners.

20. The Learning Loop

The resulting system can be understood as a continuous loop:

Observe → Interpret → Assist → Act → Measure → Adapt

Observe

What is the learner doing?

Interpret

What might the learner need?

Assist

What is the least intrusive intervention that could help?

Act

The learner makes the next decision.

Measure

Did the intervention help?

Adapt

Should the next interaction be different?

This creates a system that learns about the interaction, rather than simply answering questions.

21. A New Model of Educational AI

The EPVR experience points toward a broader model with four levels of AI assistance.

Level 1 — Conversational AI

Answers questions.

Level 2 — Contextual AI

Understands where the learner is and what they are doing.

Level 3 — Pedagogical AI

Understands how to help the learner learn rather than simply how to answer.

Level 4 — Adaptive Learning Agent

Combines conversation, environment, behavioral signals, pedagogy, personalization, and analytics.

The fourth level represents the long-term opportunity.

It is not merely an AI chatbot...It is an AI learning agent embedded within an experiential system.

22. Practical Lessons From EPVR

The development process also surfaced several practical lessons.

Latency Matters

In immersive environments, conversational timing affects the feeling of presence.

The original EPVR implementation deliberately incorporated a short delay so that the AI would not constantly interrupt the player. The project's development notes identified reducing latency as an important next step.

Voice Matters

A technically capable AI can still feel unnatural if its voice, pacing, or personality does not fit the environment.

The project identified voice quality and friendliness as areas for improvement.

Knowledge Quality Matters

A companion is only as useful as the knowledge architecture supporting it.

In educational applications, knowledge quality must be treated as a product feature—not merely a model feature.

Restraint Matters

The AI should not constantly talk.

Silence can be an important part of good interaction design.

Context Matters

The best assistance is often related to what the learner is currently doing.

23. What This Means for Training and Education

The broader opportunity is to create AI-native training environments.

Instead of taking an existing VR simulation and adding a chatbot, organizations can design the entire learning experience around an intelligent companion.

A future training environment could know:

• What the learner is attempting

• What they already know

• Where they are struggling

• What assistance they have received

• How they prefer to receive information

• Which skills they have demonstrated

• When they are ready for increased difficulty

The AI can then become an individualized instructor operating inside a shared curriculum.

This makes it possible to combine the scalability of software with some of the responsiveness traditionally associated with one-on-one instruction.

24. Measuring Success

A serious AI education platform should not measure success solely by how impressive the AI appears.

Important measurements include:

Learning Outcomes

• Knowledge acquisition

• Knowledge retention

• Transfer of knowledge

• Ability to solve novel problems

Performance

• Task completion

• Error rate

• Time to competency

• Independent performance

Engagement

• Session duration

• Voluntary exploration

• Persistence

• Interaction quality

Assistance

• Number of hints requested

• Type of assistance required

• Success following assistance

• Reduction in assistance over time

Accessibility

• Ability to customize experience

• Completion across different learner needs

• Reduced barriers to participation

AI Quality

• Factual accuracy

• Appropriate uncertainty

• Relevance

• Response timing

• Pedagogical appropriateness

The ultimate goal should be learner performance and learning, not AI usage.

25. The Human Remains Central

The most effective vision for AI in education is not an AI replacing educators, it is an AI extending what educators can accomplish.

A teacher cannot physically stand beside every learner at every moment.

An instructor cannot answer every question simultaneously.

A trainer cannot observe every interaction in a complex simulation.

An AI agent can provide an additional layer of support.

The educator remains responsible for:

• Curriculum

• Learning objectives

• Assessment

• Interpretation

• Mentorship

• Human connection

• Ethical decisions

• Complex intervention

AI provides scale and responsiveness.

26. Conclusion

Europa Prime VR demonstrated that an AI companion inside an immersive environment can serve a much broader role than conventional conversational AI.

It can be:

1. A companion.

2. A tutor.

3. A knowledge interface.

4. A puzzle guide.

5. A critical-thinking facilitator.

6. An accessibility interface.

7. A contextual assistant.

8. An adaptive support system.

9. A source of scientific and historical information.

10. A gateway to future-oriented reasoning.

11. And, when carefully designed, it can become part of the learning environment itself.

The most important insight is not that AI can answer questions. We already know that it can. The more significant opportunity is that AI can understand the relationship between the learner, the environment, the task, and the learning objective.

That changes the paradigm from:

"Ask the AI."

to:

"Learn with the AI."

And ultimately:

"Let the AI help create an environment in which the learner can succeed independently."

This is the foundation for a new generation of immersive training and education: environments that are intelligent enough to respond to the learner, disciplined enough to respect factual and pedagogical boundaries, and flexible enough to meet people where they are.

Europa Prime VR represents one practical demonstration of this direction. Developed through an NSF-funded collaboration between MXTreality and TERC, the project was built around immersive STEM learning, accessibility, adaptability, and broad participation.

The next evolution is to take the same principles beyond a science-fiction game and apply them to the places where learning matters most: classrooms, laboratories, workplaces, training centers, simulations, museums, and professional environments.

The future of educational AI may not be a chatbot waiting for a question.

It may be an intelligent companion already standing beside the learner.

Appendix A — Example AI Companion Capabilities

The AI companion developed for EPVR demonstrates a range of capabilities that can be applied to immersive education and training environments. These capabilities are not isolated functions; they can work together to create a responsive and adaptive learning experience.

Conversational Interaction

The AI can communicate naturally with the learner through voice or text, providing answers, explanations, encouragement, and social interaction. This creates a conversational interface that allows learners to interact with the experience without interrupting their immersion.

Educational application: Questions and explanations can be delivered in the context of the learner's current activity.

Context-Aware Assistance

The AI can be connected to the state of the virtual environment and understand relevant aspects of what the learner is currently doing, where they are, and which task or activity they are attempting.

Educational application: Assistance can be provided when and where it is relevant rather than requiring the learner to stop and formulate a general question.

Graduated Hints and Scaffolding

The AI can provide assistance in stages, beginning with subtle prompts and progressing toward more explicit guidance when necessary.

Educational application: Learners can receive enough support to overcome an obstacle without immediately being given the answer, reinforcing problem-solving and independent thinking.

Socratic Questioning and Critical Thinking

Rather than simply providing solutions, the AI can ask questions that encourage learners to examine evidence, consider alternatives, explain their reasoning, and reach conclusions independently.

Educational application: This approach can support scientific reasoning, troubleshooting, decision-making, and analytical thinking.

Factual Knowledge and Educational Content

The AI can provide information relevant to the subject matter contained within the experience, drawing from controlled and approved knowledge sources.

Educational application: Scientific concepts, historical information, technical procedures, terminology, and other instructional content can be made available through natural conversation.

Future and Scenario-Based Reasoning

The AI can help learners explore potential future technologies and scenarios while distinguishing established knowledge from hypothesis, extrapolation, and speculation.

Educational application: Learners can explore questions about technological development, scientific possibilities, and alternative futures while practicing evidence-based reasoning.

In-Experience Guidance

The AI can help learners navigate the environment, understand objectives, identify relevant objects, and determine possible next steps.

Educational application: Learners can remain immersed in the experience instead of repeatedly stopping to consult external instructions or menus.

Puzzle and Problem-Solving Assistance

The AI can participate directly in interactive challenges by providing contextual clues, asking questions, or suggesting strategies.

Educational application: Puzzles can become opportunities for learning and reasoning rather than simply tests of whether the learner can discover a predetermined solution.

Behavioral and Interaction Analysis

When connected to appropriate VR telemetry, the system can analyze observable interaction patterns such as movement, object interaction, repeated attempts, time on task, progression, and requests for assistance.

Educational application: These signals can help identify when a learner may be struggling, disengaging, progressing rapidly, or benefiting from additional support.

These observations should be treated as behavioral indicators rather than definitive measurements of a learner's psychological or emotional state.

Adaptive Support

Information about learner interaction can be used to adjust the AI's level of assistance.

A learner who is progressing independently may receive minimal intervention, while a learner who repeatedly struggles with the same task may receive additional hints, explanations, or alternative forms of instruction.

Educational application: The same learning environment can accommodate different levels of prior knowledge, confidence, and experience without requiring completely separate instructional paths.

Accessibility and Comfort Assistance

The AI can provide a conversational interface for accessibility and comfort features, allowing learners to communicate when aspects of the experience are uncomfortable or difficult to process.

Educational application: Depending on the capabilities of the underlying platform, the AI can recommend or assist with adjustments to elements such as brightness, audio, visual effects, movement, environmental complexity, and interaction settings.

This can make accessibility more discoverable and responsive than a conventional settings menu.

Social and Emotional Support

The AI can provide encouragement, acknowledgment, and a sense of companionship during the experience. This can be particularly valuable in immersive environments where learners may otherwise feel isolated.

Educational application: Encouragement can help learners persist through difficult tasks and feel more comfortable asking questions or requesting assistance.

The AI should be viewed as a supportive interaction layer rather than a replacement for human educators, counselors, mentors, or other professional relationships.

Progress and Performance Awareness

The AI can maintain awareness of relevant learner progress within an experience, including completed objectives, previous assistance, and demonstrated proficiency.

Educational application: The system can avoid repeatedly explaining concepts the learner has already mastered while providing additional support in areas where difficulty remains.

Instructor and Learning Analytics

Aggregated interaction data can provide educators and experience designers with insight into how learners interact with the environment.

Potential areas of analysis include:

• Where learners commonly become confused

• Which tasks require the most assistance

• Which concepts generate repeated questions

• Where engagement declines

• How long learners spend on specific activities

• Which interventions are effective

• Whether learners become more independent over time

Educational application: These insights can be used to improve curriculum, instructional design, accessibility, and future versions of the training experience.

Bringing the Capabilities Together

The real value of an immersive AI companion comes from the combination of these capabilities.

A learner may begin by asking a factual question. The AI can provide an explanation, recognize that the learner is approaching a related puzzle, offer a question rather than an answer, observe the learner's subsequent attempts, recognize when additional assistance may be useful, and provide a graduated hint.

At the same time, the system can remain available as a conversational companion and provide access to appropriate accessibility and comfort adjustments.

This creates a continuous learning relationship between:

The learner → the environment → the AI → the learning task → the learner's response

Rather than functioning as a standalone chatbot, the AI becomes an intelligent layer connecting the learner to the experience itself.

Appendix B — Example Graduated Hint System

Learner: "I don't know what to do."

AI — Level 1:         "Take another look at the equipment around you."

Learner: "I still don't see it."

AI — Level 2:         "Which device seems to be receiving information rather than sending it?"

Learner: "The receiver?"

AI — Level 3: "Exactly. What do you think it needs in order to recognize the signal?"

Learner: "The correct frequency."

AI — Level 4: "Try comparing the frequency shown on the receiver with the signals available on the control panel."

The learner solves the problem.

The AI has provided assistance without taking ownership of the solution.

Appendix C — From EPVR to Enterprise Training

The same architecture can be translated into a training environment:

VR Environment
↓
Learner Actions
↓
Interaction Telemetry
↓
AI Context Engine
↓
Knowledge + Training Content
↓
Pedagogical Reasoning
↓
Adaptive Intervention
↓
Learner Action
↓
Performance Measurement
↓
Instructor Analytics

This creates a continuous, measurable learning loop.

The result is not simply a virtual instructor.

It is a responsive training environment.

Selected References and Supporting Sources

Europa Prime VR — Official Project Site (www.europaprime.org)
EPVR describes itself as a STEM-based VR experience developed through an NSF-funded project, with an emphasis on accessibility, customization, and engagement with neurodivergent and other learners.

Europa Prime VR — Project Press Release
The project describes the companion bot as a core feature alongside interactive puzzles, exploration, and accessibility-oriented design.

EPVR AI Companion Demonstration
The original demonstration describes the integration of ChatGPT-4o into the VR experience, including social interaction, contextual guidance, additional scientific knowledge, and puzzle integration.

MXTreality — EPVR Development
Project materials describe EPVR as a science-based immersive experience developed with TERC and supported by an NSF grant, with accessibility and adaptability incorporated into the design.

Here is a version that can be downlaoded as pdf https://docs.google.com/document/d/1LudEB4pKlQ4i3CAP_PFxxBpbrM5sE-mlE_dlB1n0frA/edit?usp=sharing