Skip to main content
Back to timeline
arXivSource publication:

SearchDiffusion lets a robot hunt for objects in unknown rooms: 95.6% simulated detection, eight of ten quadruped trials successful

Synopsis

SearchDiffusion feeds a learned dense utility map into diffusion-based trajectory generation so a robot plans position and heading together in unknown environments. The map combines geometric accessibility, useful visibility and semantic progress, and serves both as a denoising condition and as an objective for sampling guidance. Across 90 simulated tasks on 24 map layouts it reached 95.6% target detection at higher planning cost than the compared planners, and on a quadruped using OneMap relevance maps it succeeded in eight of ten physical trials without fine-tuning on real-robot data.

Interpretation

When a robot searches for an object, the planner now generates position trajectories that carry heading, rather than only advancing toward the boundary of the known map. Earlier classical exploration methods plan from the robot's current map toward the boundary between known free space and unexplored regions. Across 90 simulated tasks on 24 map layouts with matched initial poses and sensing conditions, SearchDiffusion detected the target in 95.6% of tasks, versus 85.6% for fine-tuned MapEx and 85.6% for nearest-frontier planning, at greater planning cost.

Trajectory generation is conditioned on a learned dense utility map that fuses geometric accessibility, useful visibility and semantic progress into a search value for each cell. The utility map is both a conditioning input to diffusion denoising and a differentiable objective for sampling guidance, which refines candidates during the final five denoising steps. Ablations with network weights fixed toggle utility and guidance: with four map hypotheses and guidance on, adding utility raised success from 82.2% to 95.6%; with guidance off, success moved from 68.9% to 65.6%, a difference whose interval crosses zero.

Sampled counterfactual map hypotheses add geometric context about unobserved regions, but runtime hypotheses bring limited extra detection benefit. With utility and guidance both on, four, one and zero hypotheses produced 86, 84 and 85 detections. Removing hypotheses cut median planning latency from 307 ms to 236 ms, about 23%, while observed success stayed similar.

Perspective

The demonstrated scope is 90 simulated tasks on 24 map layouts plus ten physical trials on one ZSL-1 quadruped; the simulated sensor has an 18-cell range with a field of view and occlusion, and the physical planner receives a 0.1 m occupancy grid from a forward 4 m sector. The authors propose next steps of improving onboard planning efficiency, incorporating the robot's physical footprint and sensing uncertainty into trajectory generation, and evaluating in a broader range of simulated and real environments.

The authors state that search need and confidence in the utility target are search heuristics, not calibrated probabilities, and that the collision-risk terms are surrogates rather than calibrated collision probabilities. Simulation intervals do not capture inference-seed variability and are unadjusted for multiple comparisons; physical success was judged by an operator, with failures attributed to false positives in the relevance maps. The authors also note that a single physical run showing heading change alongside rising map coverage does not establish that rotation is optimal or caused solely by coverage.