Call it modular-planner dressing

Call it stacked-planner dressing. A stock perception-plus-planning stack — the same ROS building blocks running inside a thousand university labs — shows up wearing the agentic AI label, and the trick is a thin coordinator that picks between standard planners and calls it agency.

The RoboNav-Arm preprint (arXiv:2607.09716) makes the gap visible, and the gap is Gazebo-sized: the framework has been evaluated only in Gazebo Classic simulation, never on a physical arm. Its three modules are a perception layer for obstacles and ground geometry, a coordinator that updates the MoveIt collision scene (the standard ROS library for arm motion planning) and tracks progress, and a planning layer that switches between three off-the-shelf sampling-based planners — RRTConnect, RRT*, and BiTRRT. None is a foundation model. The 'agentic' part is the routing among them plus a structured semantic report that tells the planner which planner to use.

The repeatable mechanism in the RoboNav-Arm preprint: a perception module emits a tagged scene — obstacle positions, geometry, and a critical-interaction-zone label (inside, outside, within) — and the coordinator routes the goal to whichever standard planner matches the regime. Trajectory post-analysis then guarantees a collision-free path. The architecture is real; the agency claim is a costume that fits the same body.

The trap is taking 'agentic AI for robot arms' at face value. The label will harden long before the stack changes, and the next half-dozen preprints will use the same costume. The filter is mechanical: read the architecture, and if the brain is a switchboard and not a model, the label is dressing.