Navigation Among Movable Obstacles (NAMO) requires a robot to rearrange objects that block its path to a goal. For a pushing robot, the most useful action may be a setup push that does not immediately clear a passage but enables a later one. Evaluating candidate push sequences with physics simulation is costly because the number of possible sequences grows exponentially with length.
The authors introduce SPAR, a transformer-based ranker trained on simulated search outcomes. It scores push primitives using contact tokens conditioned on scene geometry and reachability regions, learning to tell immediate openings from setup pushes. The same model can guide simulator evaluations in SPAR Search or select actions directly in SPAR Policy, and both modes plug into a larger NAMO planner.
In 300 simulated two-push keyhole problems, SPAR Search reduced median simulator calls by 7.0x and planning time by 3.8x relative to Random Search. On harder problems with no two-push solution, it achieved 50.7% success within 100 simulator calls at a depth limit of five pushes, versus 20.2% for Random Search. On a real robot, SPAR Policy completed 37 of 40 trials without a physics-based digital twin, compared with 33 for SPAR Search and 27 for Random Search.