Autonomous Robotic Systems using Reinforcement Learning for Next-Gen Computing Applications

Abstract

Merger next-generation computing paradigms including edge artificial intelligence (AI) architecture, distributed sensor networks, real-time data analytics, and adaptive control systems have created a new horizons in the intelligent and autonomous robotic systems development. The above developments require having robotic agents that can undertake little supervision by humans, learn in dynamic and unpredictable environments, and make decisions in an intelligent state to realize complex goals. In this paper, we present a general frame-working approach to design, learning and deployment of Autonomous Robotic Systems (ARS) using the Reinforcement Learning (RL) in a bid to develop cognitive and operational potential of RL in the real world. Our method combines model-free and model-based RL methods which allow robots to execute tasks that concern navigation, manipulation and target tracking via constant action-feedback with the environment. As learning algorithm, we use Proximal Policy Optimization (PPO), because of its balancing between policy robustness and learning efficiency. The system is being estimated and tried out in virtual as well as real robot stages with innovations such as OpenAI Gym, Robotic operating system(ROS) and Gazebo among others. Extensive experiments show the major increases of task success rate, trajectory optimization, and resource efficiency. In particular, our trained agents using RL are up to 38 percent faster on the task completion, 27 percent less consuming energy, and are also better equipped in performing in a given scenario that has not been witnessed before as compared to conventional anticipation-based systems. What is more, we consider advanced learning methods including multi-agent reinforcement learning (MARL), curriculum learning, and continual learning to enable scalable applications in industrial automation, healthcare robotics, an urban mobility. The outcomes support the feasibility of RL-based ARS as an underpinning block of next-generation intelligent systems to provide a route with resilient, adaptable, and context-aware operations of robots in intricate, real-world settings and situations.