Human-In-The-Loop Reinforcement Learning in Autonomous Vehicles
Published:
Human-In-The-Loop Reinforcement Learning (HITL-RL) in Autonomous Vehicles (AVs) can be realized through reward shaping, human-initiated actions, and interactive learning to guide the AVs in critical situations. This can improve the efficiency of tasks and the learning ability and performance of the AVs. In addition, humans and AVs can interact and cooperate to solve tasks. In this regard, it should be noted that long-term monitoring and guidance can be physically and mentally demanding for human participants in the HITL-RL. Further, current HITL-RL methods often require expert-level demonstrations to ensure data quality and optimize performance improvement. However, expert guidance is expensive for largescale real-world applications. The main benefit of HITL-RL for AVs is to promote public trust and enhance safety.
Real-time human-UAV interactions using wearable gloves and Machine Lear (ML) based methods showcase innovative technology’s potential to enhance mission planning and control. Human and Unmanned Aerial Vehicle (UAV) interactions can be described using meta-models in emergency response scenarios to highlight the importance of formalizing human intervention points to ensure successful mission execution. HITL-RL frameworks are versatile and can be adapted to various applications, from Autonomous Driving (AD) to underwater exploration and emergency response. This versatility underlines the broad applicability and potential impact of HITL approaches in enhancing AVs across different domains. Incorporating mechanisms for human behavior modeling reduces the burden on human operators, making the system more efficient and user-friendly.
Developing a framework that learns reward functions from human driving data can effectively manage the learning process for AD, ensuring the RL agent optimizes complex goals. This approach includes trajectory sampling, offline preference learning, and RL integration to model human decision-making and obtain reward functions. Incorporating human prior experience into the reward function and safety constraints helps guide AVs in complex scenarios, such as mandatory lane changes, and limits unsafe exploration during training. Enhancing HITL-RL methods by having humans supervise, intervene, and demonstrate during the learning process not only improves the safety of AVs but also optimizes traffic flow efficiency in mixed-traffic environments. Creating a value-based HITL-DRL algorithm to address high-level decision-making problems, such as multiobjective lane-change decisions, improves training efficiency and adaptability. This approach values human policy over the undertrained DRL policy, allowing the RL agent to mimic human behaviors more efficiently. Enhancing RL performance by using human-guided learning objectives, prioritized human experience replay, and human intervention-based reward shaping improves the navigation capabilities of UGVs.
