Smooth-Mamba DRL Models Pedestrian Crash Avoidance by Vehicle Type
A recent study published on arXiv (2605.28552) presents SMamba-DDPG, a framework known as Smooth-Mamba Deep Deterministic Policy Gradient, designed to simulate pedestrian crash avoidance behaviors tailored to various vehicle types during critical safety interactions. Utilizing the Argoverse 2 dataset, the researchers analyzed genuine encounters between automated vehicles (AVs) and human-driven vehicles (HDVs). The framework develops distinct policies for pedestrian interactions with both AVs and HDVs, incorporating smooth action constraints alongside temporal representation learning. Findings indicate that SMamba-DDPG surpasses traditional reinforcement learning and supervised learning models in accurately forecasting pedestrian reactions. This research aims to improve the safe implementation of automated driving technologies by recognizing behavioral variations toward different vehicle types.
Key facts
- SMamba-DDPG framework models pedestrian crash avoidance behavior by vehicle type.
- Argoverse 2 dataset used for safety-critical pedestrian-vehicle interactions.
- Separate policies trained for AV and HDV encounters.
- Integrates smooth action constraints with temporal representation learning.
- Outperforms baseline RL and supervised learning models.
- Study published on arXiv with ID 2605.28552.
- Focuses on real-world crash avoidance behaviors.
- Aims to improve automated vehicle safety deployment.
Entities
Institutions
- arXiv