Active Sensing and Deferred-Decision Trajectory Optimization for Robust Target Identification

Autonomous systems often operate in environments where the true target is unknown among a set of candidates. A key challenge is balancing the need to gather information to identify the target with the physical requirement of remaining reachable to all potential candidates until the identity is confirmed. Existing methods like Deferred-Decision Trajectory Optimization (DDTO) prioritize reachability but are "information-passive," meaning they do not actively maneuver to improve the quality of incoming measurements. This paper addresses this gap by integrating active sensing into the DDTO framework.

The authors introduce AS-DDTO, which extends the standard DDTO objective by adding a trajectory-dependent information-acquisition term. This term biases the shared portion of the trajectories toward regions where sensing is more informative (e.g., closer to candidates). The framework supports two sensing models: a parametric Gaussian model for Bayesian belief updates and a distribution-free model using conformal prediction. To ensure the problem remains tractable, the authors derive a mixed-integer conic reformulation, allowing for efficient online replanning while maintaining theoretical guarantees on recursive feasibility and target identification.

The study demonstrates that by treating the trajectory as an active sensing control, the system can resolve target ambiguity significantly faster than standard DDTO. The inclusion of the information-acquisition term allows the planner to prioritize paths that maximize the signal-to-noise ratio of measurements. Theoretical analysis confirms that the framework maintains recursive feasibility—ensuring the system never enters a state from which it cannot reach the true target—and provides finite-sample coverage guarantees for the distribution-free candidate set.