Autonomy through humans and autonomous agents becomes more prevalent in many complex domains, including time-sensitive and unknown environments. Examples include crisis response or operational planning, where partial knowledge about casualties, locations, and the number of victims in disaster zones might be available. Several approaches have been developed to tackle the issue arising from the partial knowledge of the environment by establishing communication among agents and humans. However, communication might be limited or non-existent in complex domains with no access to communication tools or no time to process information or respond to queries. This paper develops a perception learning approach that allows agents to implicitly reason about humans’ perception of the environment using limited human data without direct communication. Human is modeled as a non-optimal reinforcement learning agent in a partially known Markov decision process. A recursive method is derived to optimally build a probabilistic model of the environment using agents’ experience and quantified humans’ perception. We demonstrate that the learned perception models can be incorporated into various decision-making policies relying on the environment model. The performance of the proposed method is investigated using a rescue operation team consisting of a human and an agent.