This paper discusses the recent progress in AI- and ML technologies for autonomous planetary surface multi-robot systems (PMRSs) and presents new concepts for multi-level robot decision and learning autonomy. Novel approaches to world model building involving mutual learning phenomena in robot teams, specifically multi-agent model-based reinforcement learning (MA/MBRL) will boost anticipatory coordination and efficient cooperation of PMRSs. MBRL will be combined with other learning approaches and information exchange protocols, such as environment map sharing and merging. The proposed techniques enable effective applications of proactive features in robot task planning, taking into account the accrued risks, time and distance travelled, as well as expected scientific benefits. Applications of the above models will focus on exploring the surface and potential near-surface liquid water reservoirs of icy moons, in particular Jupiter’s moon Europa. An example will shed light on practical techniques that can be used for learning and exchanging world models when navigating the rugged terrain of an icy moon. In the conclusion, we will argue that a multi-rover mission can harness self-supervised and multi-agent machine learning, specifically the reinforcement and federated learning, to ensuring mission resilience to threats occurring in unknown environments.