DescriptionMars exploration may lead to a better understanding of the origin of life on Earth, the discovery of life outside Earth, and the creation of a Martian habitat for humans. Even though the recent NASA Perseverance rover mission included the first successful extraterrestrial flight, robots and planetary rovers are still crucial for Mars exploration. Nevertheless, rovers still need to be perfected. For example, NASA’s Curiosity suffered severe wheel damage, which effective traction control software would have helped. This thesis presents a new machine learning-based traction control software for an autonomous Martian rover, which is developed using a reinforcement learning approach (done as numerical research analysis). Reinforcement learning is a data-driven training method, which rewards the desirable behavior of an agent (our controller) and punishes its undesirable behavior based on a predefined reward policy (tracking error). This approach is fitting because of the availability of real-time data acquired by the rover; these data train the traction controller for Mars’s unknown terrain. Specifically, in this thesis, given the unavailability of a fully equipped test bench, using numerical simulations and a range of randomized soil parameters, a Martian rover numerical model generates the data which train the reinforcement learning controller. For each wheel of the rover model used in this work, three components are necessary: a one-wheel vehicle model, a DC motor model, and a deformable soil-rigid tire model. The reinforcement learning controller used here is based on a deep deterministic policy gradient algorithm, where the controller (which outputs the motor torque) is trained to track the slip ratio. Numerical simulations in MATLAB and Simulink environments are used to evaluate the effectiveness of the proposed offline-based reinforcement learning traction control software. In conclusion, this approach advances the design of more effective planetary rovers for Mars exploration.