AI-Driven VR Disaster Resilience in Japan and India

​

Rain-induced landslides repeatedly cut off mountain roads in both Japan and India, yet accident rates have not fallen despite improved infrastructure and warning systems. The remaining bottleneck is the human driver, who must decide under uncertainty, poor visibility, and imperfect warnings.
This project, a joint effort between IIT Mandi and the Gonsalves AI Lab at Sophia University, builds a Unity-based VR driving simulator reproducing Himalayan and Hokkaido mountain roads. Warning-pole accuracy (95% vs. 70%), rainfall intensity, and time of day are systematically manipulated in 2×2 factorial experiments. Licensed drivers from both countries drive the same scenarios while steering, braking, and speed telemetry are recorded at 60 Hz alongside EEG and heart-rate variability, synchronized via LabStreamingLayer.
The resulting trajectories seed stress-aware autonomous agents: behavior cloning warm-starts the policy, which is then refined with PPO, SAC, and DQN using reward functions that balance collision avoidance against the driver's estimated cognitive and emotional load.
Outputs include an open JP–IN landslide driving dataset (telemetry + biosignals), a reusable VR simulation platform, reinforcement-learning benchmarks, and design guidance for culturally adaptive driver-assistance systems in disaster-prone terrain.