Abstract:

In this article, we delve into the practicality of using a cutting-edge deep reinforcement learning technique called proximal policy optimization. This technique is designed for tasks involving continuous control and shows promise for a specific challenge: guiding an autonomous surface vehicle that has limited control capabilities. The objective is to make the vehicle follow a predetermined path while also ensuring it avoids collisions with stationary obstacles it encounters along the route. To tackle this dual challenge, an artificial intelligence agent is employed. This agent is equipped with multiple rangefinder sensors that help it detect obstacles in its vicinity. The agent's training and performance assessment take place within a complex simulation environment, which poses various challenges. The simulation environment is generated with stochastic elements, adding an element of unpredictability to the scenarios the agent faces. The foundation of this environment is the OpenAI gym Python toolkit, which provides a platform for creating and testing AI algorithms in diverse scenarios. An interesting aspect of this approach is that the AI agent is given real-time access to its own reward mechanism. This means that as the agent operates, it can understand how its actions align with the overarching goals. Consequently, the agent has the ability to dynamically adjust its decision-making strategy. Depending on the situation, the agent can shift its focus between strictly adhering to the intended path and prioritizing obstacle avoidance to a greater extent. Upon thorough training and refining its strategies, the AI agent manages to achieve an impressively high success rate in completing its tasks. In episodic scenarios, where the agent has to follow the path and avoid obstacles, it approaches a success rate of nearly 100%. This accomplishment highlights the potential of applying advanced deep reinforcement learning techniques to complex real-world challenges in autonomous navigation and control

Description:

Proximal policy optimization, Deep reinforcement learning, Autonomous surface vehicle, Obstacle avoidance, Continuous control, Simulation environment