The Robot Reset Its Own Test Bench

HALTER builds a spatial scene graph, scores a completed manipulation rollout and plans a reset from a library of atomic skills. On four long-horizon tasks with a Franka arm, it restored the scene in 76% of episodes, versus 52% for AutoEval and 65% for a motion-planning reset. Its completed-skill estimate was correct in 90% of episodes, and reset verification in 91%; operator time fell 72% versus manual reset. On three held-out tasks, composed reset skills succeeded in 74.7% of episodes, while a per-task reset policy managed 1.3%.