Mars represents a pivotal objective in humanity’s deep space exploration endeavors, and deploying rovers for post-landing surveys is a proven strategy to extend the reach of these exploratory efforts. The Martian terrain is inherently intricate with primitive geology, bereft of prior empirical understanding, making it challenging to discern potentially perilous flat regions. The dynamic contact mechanics between rover wheels and Martian soil during traversal are difficult to accurately determine and control, inevitably leading to anomalies such as wheel embedding. When a rover encounters such sinking situations, conventional extrication techniques heavily depend on ground-controlled maneuvers, where decisions are made based on human expertise and simulated reasoning tailored to different scenarios of rover extrication. However, these approaches typically exhibit low success rates, protracted timeframes, and elevated extrication costs. In this paper, we present a novel approach through the design of a Mars rover’s autonomous and intelligent extrication control system rooted in deep reinforcement learning. This system allows the rover, following an instance of wheel extrication, to autonomously make decisions regarding speed adjustments and steering maneuvers necessary for self-extrication, utilizing its own status data. The results from simulation tests confirm that the proposed method enables the rover to successfully extricate itself under varying conditions of sinking, thus substantiating both the feasibility and practical effectiveness of the autonomous and intelligent extrication control method developed herein.