Event

SPAR: Using Search-Trained Rankers to Improve Push Action Search Efficiency

1 reports · 1 sources

Dhruv Metha Ramesh et al. propose SPAR (Search-trained Push Action Ranker), which learns from simulated search results to rank push primitives, alleviating the problem in Navigation Among Movable Obstacles (NAMO) where the number of candidate action sequences grows exponentially with length. In 300 successful runs of simulated two-push keyhole problems, SPAR Search reduced median simulator calls by 7.0x and planning time by 3.8x compared with Random Search; the same ranker can be used without retraining on 600 keyhole problems that are unsolvable within two pushes. In real-robot experiments, SPAR Policy succeeded 37/40 times without a physical digital twin, SPAR Search 33/40, and Random Search 27/40.