DreamDrive: Proactive Multi-Future Prediction for Safe Autonomous Navigation in Chaotic Mixed Traffic

This preprint presents DreamDrive, a proposed deep learning framework for proactive multi-future trajectory prediction and safe autonomous navigation in chaotic mixed-traffic environments. Instead of relying on a single most-likely future trajectory, the framework generates multiple diverse and plausible future trajectories for surrounding traffic agents. A proactive risk-aware planning layer evaluates candidate autonomous-vehicle actions against the predicted future set and explicitly accounts for potentially unsafe outcomes. The proposed trajectory decoder models multiple behavioral modes and incorporates a diversity objective to balance prediction accuracy and behavioral coverage. The framework is designed for evaluation using trajectory prediction metrics such as minADE and minFDE, together with simulation-based safety measures including collision and near-miss rates, compared against single-trajectory prediction and planning baselines.