Adaptive Robotics and Autonomous Manipulation for Flexible Manufacturing: Reinforcement Learning, Sim-to-Real Transfer, and Dexterous Manipulation in Contact-Rich Environments

Abstract

The transition from rigid, pre-programmed industrial automation to flexible, adaptive robotic systems capable of operating in unstructured and dynamic manufacturing environments represents one of the most significant challenges and opportunities in modern manufacturing. Contact-rich manipulation tasks—insertion, peg-in-hole assembly, polishing, cable routing, and deformable object handling—are ubiquitous in manufacturing but remain among the most difficult to automate, requiring the robot to reason about physical contact, comply with environmental constraints, and adapt to variations in part geometry, friction, and placement. Conventional robotic programming approaches—based on pre-specified trajectories and force thresholds—are fundamentally inadequate for these tasks: they cannot handle the variability inherent in real manufacturing environments, require extensive manual tuning for each task and product variant, and cannot adapt when conditions change. This review provides a comprehensive synthesis of adaptive robotics and autonomous manipulation for flexible manufacturing, examining deep reinforcement learning for contact-rich manipulation, sim-to-real transfer for robust policy deployment, dexterous multi-fingered manipulation, deformable object handling, and LLM-based task planning for robotic assembly. We further connect these advances to industrial sensing technologies—precision 3D optical metrology and four-dimensional thermal imaging—demonstrating their roles as enabling perceptual modalities for adaptive manipulation systems. A central contribution is the articulation of a unified Adaptive Manipulation Stack framework that integrates perception, task planning, learned control policies, and real-time adaptation for next-generation flexible manufacturing automation.