Learning to Propose Joint Sensing Actions under Shared Uncertainty

A sensing action can be valuable because of what it enables another action to reveal. Under shared receiver uncertainty, calibration can make a subsequent target observation informative even when calibration alone leaves target uncertainty unchanged. We study learning to propose executable sensing groups for a budget-limited physics-based evaluator. A two-slot Gaussian construction shows that exact singleton screening and sequential marginal greedy selection can recover an arbitrarily small fraction of the optimal variance reduction. An explicit calibration–repetition boundary identifies when resolving receiver ambiguity is more valuable than averaging extraction noise. Our recurrent proposer combines a joint target–receiver belief, learned control generation, executable projection, and group-cost prediction, with forecast feedback supporting within-decision refinement. On 192 simulated test episodes across eight scenarios, averaged over three training seeds, the full relational variant reduces mean GOSPA by 15.0% and RF energy by 7.0% relative to numerical search under matched forecast and construction budgets. Target recall improves by 4.37 percentage points, with both tracking metrics improving in every scenario mean. Fixed-checkpoint ablations identify control generation as the largest contributor: learning changes the executable choices available to the evaluator, as well as their ordering.