Year: 2024
Type: article
Abstract: Deep Reinforcement Learning (DRL) has shown immense potential in enabling robots to learn complex tasks through trial and error, without explicit programming. However, deploying DRL in real-world robotic systems presents significant challenges, including sample inefficiency, safety constraints, and the transferability of simulated results to physical environments. This paper explores the key obstacles in optimizing DRL for real-world robotics and proposes solutions to address these issues. We examine advancements in sim-to-real transfer, efficient exploration strategies, and safety-aware learning to enhance the applicability of DRL in robotics. Additionally, we discuss the integration of model-based and model-free DRL approaches to improve learning speed and performance. Through experimental evaluation on robotic manipulation and navigation tasks, we demonstrate that optimizing DRL for real-world scenarios requires a combination of robust simulation environments, hybrid learning architectures, and careful management of exploration-exploitation trade-offs. Our findings provide valuable insights for deploying DRL in autonomous systems, industrial robotics, and human-robot collaboration, ultimately bridging the gap between simulation and real-world performance.
Author: Sharma, Rahul
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7220334817