Active-search behavior + dynamic detection vocabulary when a target is unresolved

[adharshvenkat]

Problem

The semantic map is only ever populated by passive detection while driving, using a fixed launch-time prompt list (DEFAULT_PROMPTS in semantic_detector.py). Two compounding limitations:

1. Static detection vocabulary - the GroundingDINO prompt list is a launch-time ROS parameter, not derived from the task command, so the detector can never be asked to look for something outside its preset list - even though GroundingDINO itself supports arbitrary open-vocabulary prompts.
2. No active-search behavior - if a target isn't already in the semantic map (never observed, or outside the prompt vocabulary), there's no fallback exploration. The planner's own empty-plan fallback message literally says "no perceived labels yet, explore first" (llm_planner.py), but no explore behavior is implemented anywhere - it's just a log/summary string.

Proposed design

• Feed the task command's parsed target words into the detector's prompt list at runtime (dynamic vocabulary), rather than relying solely on the static launch-time list.
• Add an active-search / explore task type the planner can emit when a target is genuinely unresolved (also suggested in Planner substitutes a wrong known label instead of reporting the target unresolvable #8) - drive the robot through unexplored/likely areas looking for the target before giving up.

Dependency

Blocked on #8: the planner doesn't reliably report "unresolvable" today (it can substitute a wrong known label instead), so search can't be triggered correctly until that's fixed first.

Related

• Planner substitutes a wrong known label instead of reporting the target unresolvable #8 - planner doesn't reliably detect/report unresolved targets today
• Visualize the live semantic map in RViz via MarkerArray #9 - MarkerArray visualization would help debug search behavior once implemented

Acceptance criteria

• A command referencing an object outside the current semantic map and outside the static prompt list triggers an active search rather than an immediate failure or wrong-label substitution.
• Detection vocabulary can be extended at runtime from the task command's target labels.
• Behavior validated with at least one scripted repro (target genuinely never seen before).