Learning Occlusion-aware Decision-making from Agent Interaction via Active Perception

TL;DR

The paper tackles occlusion-induced uncertainty in autonomous driving by combining a vectorized representation of occluded environments with high-level semantic motion primitives (SMPs) and a safety-aware RL loop. It integrates a prediction model to constrain exploration within RSS-based safety boundaries, enabling risk-aware learning with improved sample efficiency. The approach, validated in challenging dynamic and static occlusion scenarios in CARLA, outperforms strong baselines (EMP, RSA, SOAP) in success rate, speed, and collision reduction while achieving real-time planning speeds. Ablation studies confirm the effectiveness of vectorized occlusion representations, SMPs, and the safety-prediction mechanism. Overall, Pad-AI advances occlusion-aware decision-making by delivering scalable, efficient, and safer active perception for autonomous driving, with potential for real-world deployment.

Abstract

Occlusion-aware decision-making is essential in autonomous driving due to the high uncertainty of various occlusions. Recent occlusion-aware decision-making methods encounter issues such as high computational complexity, scenario scalability challenges, or reliance on limited expert data. Benefiting from automatically generating data by exploration randomization, we uncover that reinforcement learning (RL) may show promise in occlusion-aware decision-making. However, previous occlusion-aware RL faces challenges in expanding to various dynamic and static occlusion scenarios, low learning efficiency, and lack of predictive ability. To address these issues, we introduce Pad-AI, a self-reinforcing framework to learn occlusion-aware decision-making through active perception. Pad-AI utilizes vectorized representation to represent occluded environments efficiently and learns over the semantic motion primitives to focus on high-level active perception exploration. Furthermore, Pad-AI integrates prediction and RL within a unified framework to provide risk-aware learning and security guarantees. Our framework was tested in challenging scenarios under both dynamic and static occlusions and demonstrated efficient and general perception-aware exploration performance to other strong baselines in closed-loop evaluations.

Figures (9)

• Figure 1: (a) Common occlusion scenarios. (b) Instead of conservatively following the slow truck, our method will cautiously reduce the occlusion uncertainty with an active probe motion. Based on updated observations, our method swiftly decides whether to overtake or abort the overtaking.
• Figure 2: Schematic architecture: Vectorized observations from the occluded environment are encoded through a graph neural network. The actor decodes parameterized action probabilities, which are further mapped to the semantic motion primitives (SMPs). A safe interaction mechanism is engaged to avoid risky motion primitive exploration and security guarantees.
• Figure 3: Vectorized representation of occluded environments, including visible region, lane lines, and agent history.
• Figure 4: Safety Interaction Mechanism
• Figure 5: RSA driver continuously follows the slow truck, limiting the ability to change lanes actively.