TL;DR
ProACT is a performance-driven adaptive training system for VR firefighting that dynamically adjusts fire behavior to keep trainees in the "Goldilocks zone" of challenge. By combining real-time control theory for global fire spread with LLM-based semantic reasoning for local flammability, the system ensures training remains engaging and effective regardless of a user's skill level.
What Problem Does This Paper Address?
Firefighter training is a high-stakes endeavor. Real-world live-fire exercises are prohibitively expensive, dangerous, and difficult to scale. While Virtual Reality (VR) offers a safe alternative, most current systems suffer from a "static difficulty" problem: scenarios are either too easy for veterans or overwhelmingly hard for novices.
Effective skill acquisition requires maintaining an appropriate challenge level—aligning task difficulty with individual proficiency. However, adjusting difficulty in a complex, dynamic environment like a fire is non-trivial. It requires balancing three competing needs:
- Real-time Regulation: Responding instantly to trainee actions.
- Precise Regulation: Adjusting specific environmental factors (like material flammability) rather than just "turning up the heat" globally.
- Stability: Ensuring the fire behaves realistically and doesn't change in jarring, immersion-breaking ways.
Methods
The authors introduce ProACT, which treats the simulated environment itself as the adaptive variable. The system architecture (illustrated in Figure 2) uses a multi-scale approach:
- Global Fire Dynamics Regulation Module: This operates on a fast timescale (every 0.5s). It uses a feedback structure inspired by Proportional-Integral (PI) control. It calculates an "experience gap" by comparing the trainee's actual fire suppression rate against a target trajectory. If the trainee is too fast, it increases the global fire spread probability; if they are struggling, it slows it down.
- Local Semantic Scene Reasoning Module: This operates on a slower timescale (every 20s). It uses a Large Language Model (GPT-5-mini) to perform "semantic grounding." The LLM analyzes a "Static Library" (object materials like wood vs. metal) and a "Dynamic Library" (trainee movement history) to identify specific regions where flammability should be adjusted. This ensures that if the fire needs to get harder, it does so by making logical objects (like a wooden chair) more flammable, rather than just increasing spread in empty space.
Datasets
The researchers evaluated ProACT using three distinct setups rather than standard external datasets:
- 2D Grid-World Simulation: An abstract environment used to isolate and test the global challenge regulation logic against simulated agents with varying "error rates" ($\epsilon$-greedy models).
- 3D Object-Aware Simulation: A more complex environment with realistic geometry and heterogeneous flammability to test the coordination between global and local modules.
- VR User Study: A study involving 8 participants (ranging from beginners to experts) to evaluate subjective experience and objective performance in a high-fidelity VR kitchen fire scenario.
Key Insights and Ideas
- Environment as the Controller: Most adaptive VR focuses on changing the UI or the instructions. ProACT changes the physics of the hazard. By making the environment part of the feedback loop, the training objective remains unchanged while the difficulty scales naturally.
- The Multi-Scale Hybrid: The "aha moment" here is the division of labor. Numerical control (PI) handles the "how much" of difficulty adjustment with high responsiveness, while the LLM handles the "where" and "why," ensuring the changes make semantic sense to the human trainee.
- Avoiding Explicit User Modeling: Instead of trying to guess a user's "skill level" or "intent"—which is notoriously difficult in VR—ProACT reacts directly to observable environmental outcomes (fire load and suppression pace). This makes the system robust to different styles of play or temporary distractions.
Key Results
ProACT was compared against three baselines: StaticEasy, StaticHard, and a simple Reactive system.
- On-Target Completion: In 3D simulations, ProACT achieved significantly higher "on-target" completion rates (finishing within the 60-120s window) across all skill levels. For the most inconsistent trainees, ProACT hit the target 70% of the time, compared to just 23% for the Reactive baseline.
- User Preference: In the VR study, ProACT received the highest mean ratings across all eight subjective dimensions, including engagement, realism, trust, and challenge appropriateness (see Figure 10).
- Pacing Stability: As shown in Figure 11, ProACT successfully managed the fire load to ensure it didn't grow out of control (like StaticHard) or vanish too quickly (like StaticEasy), creating a consistent "burn" that maximized practice time.
Limitations and Future Work
- Collaborative Training: Currently, ProACT is designed for solo trainees. Real firefighting is a team sport. Future work needs to explore how to synchronize difficulty adaptation across multiple users in a shared space.
- Human Decision Modeling: The simulation used simplified $\epsilon$-greedy agents. Real humans exhibit complex stress responses and navigation strategies that the current simulation doesn't fully capture.
- Scale of Study: The user study was small ($n=8$). Larger, more diverse populations (including professional firefighters) are needed to validate the system's training effectiveness.
- Generalization: While focused on fire, the authors suggest this "environment-as-feedback" approach could apply to other domains like disaster evacuation or industrial safety training.
Summary updated