Autonomous systems get a live gauge for information gathering: AST separates idle looping from purposeful exploration before the task is done
Synopsis
Autonomous systems usually monitor motion and whether a task is complete, leaving the information-gathering process itself unchecked, so a system can keep operating without resolving outstanding requirements. Acquisition-State Telemetry (AST) maps raw observations onto task-requirement resolution coordinates, and a fixed causal monitor emits two continuous signals: prospective progress across unresolved requirements and informational friction when recurrent evidence yields negligible gain. One unretuned configuration ran across eight domains; the friction channel separated unproductive stagnation from purposeful exploration before completion and stayed muted afterward, and in closed-loop reinforcement learning under early termination, penalizing friction sustained high mission success.
Interpretation
Autonomous systems gain a runtime readout of whether information gathering is advancing declared task requirements, not merely whether the system is still active. Prior monitoring covered physical kinematics, environmental states and binary task completion, while information acquisition itself went unmonitored at runtime, so a system could stay continuously active without resolving any outstanding requirement. A single unadjusted monitor configuration ran across eight domains spanning search-and-rescue dialogue, spatial grid exploration, capability unlocking, continuous robotic manipulation and probabilistic intent estimation; the friction channel distinguished unproductive stagnation from purposeful exploration before completion and remained muted during post-completion activity.
Two signals answer different questions: one estimates expected progress on unresolved requirements, the other rises when repeated evidence adds almost no new information. The monitor uses declared task requirements rather than physical dynamics as its fixed reference frame, and a domain interface converts raw observations into cumulative resolution per requirement plus normalized representations for comparison. Friction measures recurrence as the orthogonal complement of historical evidence representations, weighted by realized gain and scaled by each requirement's unresolved fraction; in the controlled search-and-rescue interviews, efficient questioning resolved all 8 categories with mean friction 0.02, while the two recurrent regimes stopped near half completion with mean friction 0.26 and 0.28.
Putting friction into the reward lets agents complete most mission requirements when the interaction budget is cut short. Telemetry in observations alone, or an operational-cost penalty matched to the friction penalty's magnitude, did not escape repetitive questioning loops. Among PPO policies trained on search-and-rescue corpora under early termination, the unpenalized baseline resolved 2.37 categories with a 7% full-resolution rate, the matched operational-cost comparator resolved 3.53 categories at 20%, while friction-penalized AST-Pen resolved 7.07 categories at 87% and Full-AST resolved 6.93 at 83%, both with zero stall burden.
Perspective
The demonstrated scope is one fixed monitor configuration across eight heterogeneous domains, including controlled search-and-rescue interviews, MultiWOZ, Map Task, AnnoMI, MiniGrid, Crafter, MetaWorld and BOIR intent estimation, plus closed-loop PPO experiments on search-and-rescue corpora; the monitor assumes a finite, declared set of task requirements that keeps consistent definitions within a monitoring episode.
Monotonic accumulation assumes previously verified information stays valid until an explicit episode reset, so transient observations, time-varying facts or active misinformation would require extending the state contract beyond monotonic updates; the absolute magnitude of telemetry depends on the resolution mapping and representation space the host system supplies, and replacing expert scoring with a five-model language-model ensemble preserved the friction ordering across regimes while changing absolute values.