Is the Agent Aware of the Situation? Probing Embodied DRL Agents' Situation Assessment

Understanding and evaluating the behavior and correctness of agentic systems requires more than analyzing agents' externally observable behavior; it also requires methods for inspecting their internal task-relevant states. For context-dependent decision making, an embodied agent needs to infer the task-relevant situation it is currently facing from its various sensor measurements, since different situations may require different actions. While Deep Reinforcement Learning (DRL) is a popular building block of agentic systems for learning the control behavior of embodied agents, policies trained with DRL largely operate as black boxes. This makes it difficult to determine whether and where the embodied agent internally encodes its assessed situation to trace back if the agent correctly interpreted its environment. We therefore investigate whether the Human Factors concept of Situation Awareness can be adopted as a framework towards operationalizing and measuring a DRL agent's environment understanding. We present a controlled experiment implemented in Meta AI's Habitat 3.0 simulation platform requiring a robotic agent to learn a situational navigation task, defining controlled situation variables. Linear probes of the policy's internal layer representations on the controlled external situation provide empirical evidence that the situation is strongly encoded in the policy's final internal layer representation passed to its action selection head. These results suggest that situation-awareness probing can provide an operational mechanism for making task-relevant internal agent states more observable, offering a measurement building block for the analysis of larger agentic systems.