Source-linked AI summary
Too Hard, Too Easy, Just Right: Performance-Driven Challenge Adaptation in Virtual Reality Firefighter Training
Xuan Wang, Jianan Jiang, Rongqian Chen, Tian Lan, Bin Li
TL;DR
Firefighter VR training needs to accommodate changing trainee performance, but many systems use fixed scenarios and live-fire practice is risky and costly. ProACT adapts simulated fire behavior through continuous global regulation and periodic local flammability updates. It better maintained an appropriate challenge level and improved perceived outcomes compared with adaptive and non-adaptive baselines.
Problem
Many VR firefighter training systems rely on fixed scenarios or limited adaptation, despite the need to match challenge to changing trainee performance and the risks and costs of live-fire practice.
Source: §1 · p. 2Method
ProACT regulates simulated fire behavior using continuous, performance-driven scene-level spread control and periodic LLM-based object-level flammability updates.
Source: §1 · p. 2Results
ProACT better maintained a just-right challenge level and improved perceived challenge appropriateness, trust, and reuse intention versus adaptive and non-adaptive baselines.
Source: §8 · p. 14Takeaways & Limitations
Performance-driven environment regulation may inform immersive training beyond firefighting, including emergency response and industrial safety.
Source: §7.3 · p. 13, §8 · p. 14The exploratory VR study was modest in scale (n=8), used university-community participants, and relied on custom Likert items rather than fully standardized instruments.
Source: §7.2 · p. 13
Abstract
from arXiv · showhide
Effective firefighter training requires practice under dynamic and stressful emergency conditions, yet real-world live-fire exercises are costly, risky, and difficult to scale. Virtual Reality (VR) offers a safe and controllable alternative, but many existing VR firefighter training systems rely on fixed scenarios or provide limited difficulty adaptation, making it difficult to maintain an appropriate level of challenge for skill development as trainees' performance changes over time. To address this limitation, we present ProACT, a performance-driven adaptive training system that regulates scenario difficulty through closed-loop interactions between trainee performance and simulated fire dynamics. Specifically, ProACT combines a global module that continuously adjusts scene-level fire spread dynamics in real time based on trainee performance with a local module that periodically updates object-level flammability across the environment. Our evaluation shows that ProACT improves both objective training performance and subjective user experience compared with existing adaptive and non-adaptive baselines. We further discuss how adaptive immersive environments can extend beyond firefighting to broader training domains.
1 Introduction
VR can provide safe, controllable practice for hazardous firefighting, but fixed scenarios may not match trainees’ changing proficiency. ProACT responds by adapting fire dynamics to performance, with reported benefits in training performance and perceived challenge.
1 Introduction
- VR offers controlled exposure to dynamic, stressful firefighting conditions that are dangerous, costly, resource-intensive, and difficult to scale in live-fire training.Live-fire safety constraints also limit how frequently such exercises can be conducted.Source: §1 · pp. 1–2
- Existing systems often use fixed or instructor-adjusted scenarios, although effective skill acquisition depends on matching task difficulty to individual proficiency.Firefighting challenge also emerges from fire propagation, object ignition, spatial layout, visibility, and trainee actions.Source: §1 · p. 2
- Performance-driven firefighting adaptation remains uncommon, and systems must map changing task signals to timely, precise adjustments without making fire behavior unstable or unrealistic.The identified challenges include real-time regulation, fine-grained control of interacting environment factors, and balancing responsiveness with stability.Source: §1 · p. 2
- ProACT combines continuous, performance-driven scene-level fire-spread regulation with periodic LLM-based estimation of object-level flammability and ignition likelihood.It monitors elapsed time, extinguishing progress, and fire evolution to adapt fire behavior across spatial and temporal scales.Source: §1 · p. 2
- ProACT’s evaluation reports higher on-target completion, better adherence to target challenge ranges, and more appropriate perceived effort and challenge than existing baselines.The reported benefits also include stronger perceptions of challenge appropriateness.Source: §1 · p. 2
2 Related Work
Prior work establishes immersive training, dynamic difficulty adjustment, and context-aware AI as relevant foundations. This paper’s related-work framing highlights less explored regulation of the simulated hazard environment itself.
2.1 Virtual and Immersive Training Systems
- Immersive training spans procedural rehearsal and emergency-response simulation, while much work emphasizes guidance and feedback over dynamically regulating the hazard environment.Examples include visual and haptic feedback, corrective guidance, and multimodal interaction.Source: §2.1 · pp. 2–3
2.2 Dynamic and Adaptive VR Training
- Dynamic difficulty and adaptive VR training use performance, behavioral, or physiological signals, but prior approaches largely adapt instruction rather than the simulated environment.Firefighter VR research has focused more on feasibility, ecological validity, stress induction, and human factors than performance-driven scenario control.Source: §2.2 · p. 3
2.3 Generative AI and Context-Aware Intelligence
- Generative AI and context-aware immersive systems support interpretation and personalized interaction, but rarely use generative models to regulate training-environment dynamics.ProACT uses LLM-based semantic reasoning to estimate object-level flammability and adapt local ignition behavior.Source: §2.3 · p. 3
3 Adaptive Firefighter Training System
ProACT couples real-time global fire-spread control with periodic, semantically informed local flammability updates. Its control objective is to keep training within a target challenge range while coordinating responsiveness with stable fire behavior.
3.1 System Overview
- The system monitors environment state, trainee movement, and task performance, then coordinates global spread control with local object-level ignition updates.These complementary mechanisms address real-time difficulty regulation, fine-grained environment control, and stable but responsive adaptation.Source: §3.1 · p. 3
3.2 Environment Dynamics
- ProACT represents fire and flammability as spatial fields and defines success as completion within a target-time window, not minimizing completion time.A fast loop adjusts scene-level spread probability, while a slower loop updates spatial flammability using semantic and behavioral context.Source: §3.2.1 · p. 4, §3.2.3 · p. 5
3.3 Global Fire Dynamics Regulation Module
- The global controller compares normalized fire-load trajectory and suppression-pace errors, shifting weight from short-term pacing early to target-trajectory convergence later.Positive experience gaps reduce spread probability to alleviate overload; negative gaps increase it to preserve challenge, with smooth adjustment rather than predefined difficulty tiers.Source: §3.3.2 · pp. 5–6, §3.3.3 · p. 6
3.4 Local Semantic Scene Reasoning Module
- The local module fuses persistent object semantics with recent trainee behavior to select regions for periodic, bounded flammability updates.A consistency check constrains target regions and update magnitude to keep adaptation aligned with global challenge regulation and prevent abrupt fire changes.Source: §3.4 · p. 6, §3.4.1 · p. 7, §3.4.2 · p. 7
4 Experimental Setup
The evaluation used Unity and a Quest 3 for the VR user study, with a matching Python simulator for simulation-based evaluation. ProACT was compared with three baselines, including fixed-difficulty policies and a lightweight reactive controller.
4.1 Implementation Details
- The VR user-study system ran in Unity 2022.3.43f1 on a Meta Quest 3, while simulation used a Python implementation of the same fire-dynamics abstraction and adaptation logic.The global controller ran every 0.5 seconds in Unity and used normalized time steps in simulation.Source: §4.1 · p. 8
- The local semantic scene reasoning module updated object-level flammability every 20 seconds or steps using GPT-5-mini to guide bounded adjustments in selected regions.The implementations shared a discrete update index, though a simulation step and a Unity step represented different physical durations.Source: §4.1 · p. 8
4.2 Baselines
- ProACT was compared with StaticEasy, StaticHard, and Reactive, which respectively represent fixed under-challenging, fixed over-challenging, and coarse online feedback policies.Reactive adjusts spread probability based on recent extinguishing rate and predicted completion time relative to a target window.Source: §4.2 · p. 8
5 Simulated Environment Evaluation
Across controlled simulations, ProACT most consistently steered completion toward the target window, first in a 2D grid and then in a 3D object-aware setting. The setup varied simulated trainee capability and compared ProACT with fixed and reactive baselines.
5.1 Evaluation Setup
- The staged evaluation used a 2D grid without object semantics to isolate global regulation, then a 3D simulation with geometry and semantic flammability to assess both modules.Each setting compared four methods across four simulated capability levels, with 100 episodes per method and capability level and a target completion time of 90 steps.Source: §5 · p. 8, §5.1 · p. 9
5.2 2D Grid-World Simulation
- In the 2D grid, representative trajectories show ProACT gradually aligning fire-load reduction with the target window, while StaticEasy and StaticHard produce under-challenge and overload, respectively.Reactive’s threshold-based updates can lag behind the evolving fire state, leaving some trajectories outside the target window.Source: Figure 7 · p. 9, §5.2 · p. 10
5.3 3D Object-Aware Simulation
- In the 3D simulation, ProACT had the highest on-target rates at ε=0.05, 0.20, and 0.60 (49.0%, 48.0%, and 70.0%), including a 98.10-step mean at ε=0.60 versus 133.97 for Reactive.ProACT was strongest at the extreme capability levels; at ε=0.40, Reactive reached 66.0% OT.Source: §5.3 · p. 10
- In 3D trajectories, ProACT coordinated global spread regulation with object-aware flammability updates, completing near the target time at ε=0.60 while other methods remained outside the window or finished late.The 3D simulation incorporated scene geometry and heterogeneous object flammability, unlike the uniform-flammability 2D grid.Source: §5.3 · p. 10
6 VR User Study
An exploratory within-subjects VR study compared four fire-training conditions using task logs and post-condition ratings. Participants rated ProACT highest on every questionnaire item, while logged fire-load trajectories showed it finishing within the intended time window more consistently than the baselines.
6.1 Study Design
- Each participant experienced StaticEasy, StaticHard, Reactive, and ProACT in randomized order, with task completion time and extinguished fire cells recorded for a 60–120 s target window.Trials ended when fires were extinguished or after 150 seconds; participants also completed an eight-item questionnaire after each condition.Source: §6.1 · p. 10, Table 2 · p. 11, Figure 8 · p. 11
6.2 Participants and Ethics
- The exploratory study recruited eight university-community participants aged 23–30 with varied XR experience, and the protocol received institutional review board approval.The group included two beginners, two intermediate users, and four experts.Source: §6.2 · p. 11
6.3 Procedure
- Participants practiced navigation, fire observation, and extinguisher-based suppression without adaptive regulation before completing the four randomized formal trials.They rested for at least one minute between trials and completed questionnaires after each condition.Source: §6.3 · pp. 11–12
6.4.1 Questionnaire Results.
- ProACT received the highest mean rating on all eight questionnaire items, with scores above 5.3/7, and outscored Reactive most on trust, challenge appropriateness, and reuse intention.The items covered engagement, realism, adaptation, trust, challenge, effort, relevance, and willingness to reuse.Source: §6.4.1 · p. 12, Figure 10 · p. 12, Table 3 · p. 12
6.4.2 Log-Based Fire-Load Analysis.
- Logged trajectories showed ProACT’s fire load fluctuating with suppression progress before reaching zero within the 60–120 s target window, unlike the early-ending StaticEasy and often uncontrollable StaticHard patterns.Reactive showed better control than the static baselines, but its fire-load trajectory changed more smoothly and slowly than ProACT’s.Source: §6.4.2 · p. 12
7 Discussion
ProACT adapts challenge by regulating fire dynamics in response to trainee progress, while its evaluation and broader applicability remain bounded by simulation and study scope. The discussion frames environment-level adaptation as a potential fit for other immersive training tasks with observable progression and controllable dynamics.
7.1 Implications
- ProACT regulates challenge from observable task progress, including suppression pacing and remaining fire load, without predefined difficulty levels or explicit user-skill models.Its continuous global controller governs fire progression, while periodic semantic reasoning identifies bounded local adjustments from scene context; the LLM does not continuously control difficulty.Source: §7.1 · pp. 12–13
- ProACT generally brings active fire load to zero within the 60–120 s target window, while StaticEasy ends too early and StaticHard grows out of control.Figure 11 shows individual trials, condition means, and ±1 standard-deviation bands.Source: Figure 11 · p. 13
7.2 Limitations
- The evaluation is limited by simplified trainee behavior and controlled fire dynamics, and completion time measures task pacing rather than perceived challenge or learning effectiveness.The simulated trainee does not capture the full range of human decision-making, stress responses, or navigation strategies in real emergencies.Source: §7.2 · p. 13
- The exploratory VR study is modest in scale (n=8), uses university-community participants and custom Likert items, and lacks objective learning-transfer outcomes.The authors identify broader populations, validated measures, and learning-transfer outcomes as directions for future assessment.Source: §7.2 · p. 13
7.3 Broader Applicability
- ProACT may apply to immersive training where task progression is observable online, environment dynamics have controllable parameters, and scene context can constrain local adaptations.The discussion names emergency response, disaster evacuation, and industrial safety as domains with evolving hazards or task pressure.Source: §7.3 · p. 13
8 Conclusion
ProACT addresses fixed or limited-adaptation VR firefighter scenarios with closed-loop regulation of simulated fire behavior. Across controlled 2D and 3D simulations and a VR user study, it better maintained an appropriate challenge level and improved several subjective experience measures against adaptive and non-adaptive baselines.
8 Conclusion
- The conclusion identifies fixed scenarios or limited adaptation as a constraint in prior VR firefighter training systems for trainees with different skill levels.ProACT was presented to continuously regulate scenario difficulty based on trainee performance.Source: §8 · p. 13
- Across controlled 2D and 3D simulations and a VR user study, ProACT better maintained a “just-right” challenge level than adaptive and non-adaptive baselines.The reported subjective improvements included perceived challenge appropriateness, trust, and reuse intention.Source: §8 · p. 14
- ProACT couples a real-time global fire-dynamics controller with a periodic, LLM-based local flammability-reasoning module.The system regulates simulated fire behavior through closed-loop interactions between trainee performance and the environment.Source: §8 · pp. 13–14