Robots plan the search before moving: PreNav's diffusion generator cuts reference generation from 25.7 seconds to about one
Synopsis
In semi-static indoor scenes where walls and large furniture stay put but small objects move, PreNav stops re-querying a vision-language model mid-navigation and instead generates the whole search reference before the robot moves: a large language model scores target-furniture co-occurrence, and a conditional diffusion model turns that prior, the wall raster and the start pose into an ordered centerline with a search band around it. On the six-category Gibson ObjectNav subset with matched wall and five-category furniture priors, PreNav reaches 64.3% success and 39.2% SPL against SGM's 62.0% and 36.2%; the generator is trained only on room-level ScanNet++ scenes and never fine-tuned on Gibson.
Interpretation
The robot moves the decision of where to look and in what order out of the navigation loop and into a single global search reference computed before it starts, leaving only online target detection and approach during the run. Prior-map systems answer queries from object poses recorded at reconstruction time, which expire once objects move, while map-free methods re-derive candidate regions from current observations every few steps, often through repeated vision-language model calls. On the six-category Gibson ObjectNav subset with the same wall and five-category furniture priors, PreNav reaches 64.3% success and 39.2% SPL versus SGM's 62.0% and 36.2%; adding the scene's unlabeled furniture as 2D boxes raises PreNav's success by a further 6.0% and SPL by 0.6%.
The search reference is assembled in two parts: a large language model scores target-furniture co-occurrence offline, and a diffusion model spreads those scores over the known furniture to produce an ordered centerline with a width band. The co-occurrence scores are the generator's only query-dependent input, so a new query needs one scoring run rather than retraining the generator. The generator trains only on 83 room-level ScanNet++ scenes (75 training, 8 validation) and is applied directly to floor-level Gibson; against references from a traveling-salesman solver, diffusion generation takes 0.995 s on average and 1.029 s at P95 versus 25.705 s and 169.893 s, with 0.7% lower success and 0.2% higher SPL on the six-category subset.
Perspective
The demonstrated scope is the ObjectNav benchmark on five Gibson v1.1 validation scenes (600 episodes in the six-category subset, 1,000 across ten categories, 500-action budget) plus 15 open-vocabulary trials on a Fetch robot with a RealSense D435i and SAM 3 detector in a laboratory of over 200 square meters; the method assumes the target lies on or near furniture.
The method assumes the target is on or near furniture, so an object left on open floor falls outside the band and is not found, as happened with a quadruped robot on the lab floor; on the ten-category benchmark the full-scene furniture prior adds 0.3% success but costs 0.8% SPL, meaning slightly longer routes for a few more found objects; book, bottle, clock and cup are limited by detection rather than search, and the 3DSG annotation and the top-down semantic map used for success disagree at the category level.
