Aug 2026· Machines· Vol 14, pp. 900· 0 citations· 35 references
TL;DR
The proposed framework improves the safety and crossing efficiency of autonomous vehicle decision-making at unsignalized intersections and introduces a composite prioritized replay mechanism into the Twin Delayed Deep Deterministic Policy Gradient algorithm.
Abstract
Due to the absence of traffic signal control and the difficulty in accurately estimating the future movements of surrounding vehicles, autonomous vehicle decision-making faces challenges at unsignalized intersections. This study proposes a trajectory prediction-aided deep reinforcement learning framework. First, a composite prioritized replay mechanism is introduced into the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, jointly considering temporal-difference error and reward-based event severity to enhance critical-experience reuse. Second, a convolutional multi-layer long short-term memory (CM-LSTM) model predicts surrounding-vehicle trajectories through convolutional local-motion encoding and stacked LSTM temporal modeling, and the predicted trajectories are incorporated into the deep reinforcement learning state representation. A multi-objective reward function is designed to balance collision avoidance, passing efficiency, lane keeping, and task completion. In CARLA go-straight and left-turn tests, CLS-TD3 achieves success rates of 93.8% and 90.2%, collision rates of 2.5% and 4.2%, and average passing times of 5.18 s and 5.58 s. Compared with TD3, the success rates increase by 6.3 and 8.6 percentage points, while average passing times decrease by 18.8% and 20.5%. These results demonstrate that the proposed framework improves the safety and crossing efficiency of autonomous vehicle decision-making at unsignalized intersections.
In mixed traffic, decision-making for autonomous vehicles (AVs) confronts three interrelated challenges. First, physics-based priors incorporated into reinforcement learning (RL) models fail to capture latent interactive vehicle intentions and diverse driver behaviors, limiting the proactive reasoning capabilities. Second, abrupt maneuvers by surrounding vehicles cause non-stationarity, leaving long-tail safety events under-explored. Third, hybrid action spaces destabilize unified RL training due to the different temporal scales of continuous car-following and discrete lane-changing maneuvers. To address these issues, we propose Knowledge-Data Dual-driven Reinforcement Learning (KDDRL). First, a conditional deep generative model synthesizes intention-aware future trajectories, converting passive perception into proactive predictive states. Second, a knowledge-data dual-driven paradigm operates on these predictive states, fusing probabilistic data-driven insights with physical constraints to guide safe exploration through safety-critical scenarios. Third, a coupling module compresses both intention-aware trajectories and physical constraints into compact shared embeddings. This unified representation enables asynchronous multi-timescale optimization of continuous car-following and discrete lane-changing while preserving mutual information. Evaluations on dataset-calibrated simulations demonstrate that KDDRL effectively handles intention uncertainty, accelerates training convergence, and outperforms conventional baseline methods in terms of safety, efficiency, and comfort.
Jie Fang, Wei Zheng, Mengyun Xu et al.· 0 citations
Autonomous unmanned aerial vehicle path planning in rescue missions must cope with dynamically moving targets and obstacles, where decision-making based solely on instantaneous observations often becomes myopic and fails to anticipate future motion behaviors. To address this issue, this paper proposes a trajectory-predictive and attention-enhanced deep recurrent State-Action-Reward-State-Action (SARSA) algorithm (PA-DR-SARSA). This algorithm incorporates future motion information into on-policy reinforcement learning through a Gated Recurrent Unit–based trajectory prediction module that learns temporal motion patterns from historical observation sequences. Short-horizon trajectory forecasts are fused with current observations and attention-enhanced features to construct a prediction-enhanced decision representation for SARSA action-value evaluation. Furthermore, an attention mechanism is introduced to adaptively weight predictive and instantaneous features, enabling the agent to prioritize decision-relevant motion information. Moreover, a risk-aware reward shaping strategy leverages these predicted trajectories to guide proactive action evaluation under dynamic uncertainty. Simulation results in dynamic grid-based rescue environments demonstrate that, compared with representative planning and reinforcement learning baselines, the proposed algorithm reduces average path length by up to 13.2% and turning points by up to 39.3%, while improving dynamic obstacle avoidance and task success rates by up to 17.2% and 23.7%, respectively, without compromising on-policy learning stability.
Unknown authors· Transactions of the Institut...· 0 citations
The proposed learning-task formulations improve prediction performance across the selected benchmark datasets compared with the standard supervised-learning formulation and demonstrate that reformulating the decision process and training objective can improve an advanced pedestrian trajectory prediction architecture.
Šimon Sukup, Ariyan Bighashdel, P. Jancura· 0 citations
Efficient and reliable path planning remains a core challenge for autonomous
vehicles operating in dynamic and crowded environments. Although Deep
Reinforcement Learning (DRL) has shown considerable potential in autonomous
decision-making, it still faces challenges such as insufficient feature
extraction, sparse rewards, and low obstacle avoidance efficiency in complex
scenarios. To address these issues, this paper proposes an end-to-end path
planning framework, PPO-ICM-Attn. Built upon the Proximal Policy Optimization
(PPO) algorithm, the framework incorporates a dual-channel attention
convolutional neural network module (Attention-CNN) to enhance spatial and
semantic understanding of dynamic obstacles, and introduces an Intrinsic
Curiosity Module (ICM) to promote active exploration in sparse-reward settings.
Furthermore, a reactive avoidance reward function based on velocity-obstacle
theory is designed and embedded to achieve real-time proactive collision
avoidance in highly dynamic environments. Experiments are conducted in a
semi-structured dynamic crowd scenario constructed on the GAZEBO simulation
platform. The results demonstrate that PPO-ICM-Attn achieves significant
improvements in key metrics such as path success rate, travel time, and path
efficiency compared to baseline methods like A*+DWA and standard DRL. Although
the gap remains in path efficiency compared to A*+DWA, the proposed method
exhibits superior robustness and navigation performance overall, validating its
effectiveness in complex dynamic environments.
Shiquan Shen, Jiahao Liu, Zheng Chen et al.· SAE technical paper series· 0 citations
Autonomous vehicles (AVs) have shown significant potential in recent years, with increasing interest from the public, industry, and academia in their adoption on roads. To achieve reliable autonomous driving, AVs must be able to operate safely under adverse weather conditions such as rain-induced wet roads which pose serious safety risks to road users. Therefore, AVs require a robust Lane Change Decision (LCD) model as a crucial part of their decision-making system. Although existing studies have explored the use of deep reinforcement learning (DRL) for LCD models, most of these works do not consider rainy conditions in the development of the LCD model. To address this gap, this paper proposes a rain-aware LCD model using Deep Q-Networks (DQN). The proposed framework was implemented using the open-source
Highway-env
simulation environment, modified to incorporate rain-dependent human-driven vehicle behavior and wet-road friction effects on the ego vehicle. In addition, a safety-augmented DQN model was introduced to penalize unsafe close-following behavior before an actual collision occurs. The proposed models were evaluated in terms of safety, efficiency, comfort, and global traffic impact against rule-based (MOBIL) and learning-based benchmarks. The proposed Rain-aware DQN model improved the overall lane-changing performance by 11.2% and 17.7% according to two composite evaluation metrics, compared with MOBIL. The proposed Safety-augmented DQN model also achieved strong performance, with corresponding improvements of 10.6% and 17.3%. Moreover, comparison with other DQN models showed that the proposed Rain-aware DQN improved the safety index by 2.1% and 3.0% respectively.
A. Alzubaidi, R. Babu, Young-Ji Byon et al.· The Arabian journal for scie...· 1 citation
A vision-based end-to-end autonomous parking framework trained through imitation learning that introduces a historical context fusion encoder to capture temporal dependencies from past vehicle motions, a dual-stream attention decoder to enhance interaction between scene features and trajectory representations, and kinematic-aware auxiliary losses to enforce smooth and feasible trajectory generation.
Daisy X. M. Zheng, Bingli Zhang, Xinyu Wang et al.· Engineering Research Express· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
