ZONDA: Zero-shot Multi-floor Object Navigation with Dynamic Avoidance, Tested on Biped Robot

rsasaki0109 · x · 2026-08-14

ZONDA: Zero-shot Multi-floor Object Navigation with Dynamic Avoidance

Background: Existing object goal navigation methods are limited to static, single-floor environments, ignoring cross-floor topologies and dynamic pedestrians, hindering real-world deployment.
Method: Proposes ZONDA, integrating three core components:
Heuristic multi-floor planning: Uses height-difference traversable maps for stair traversal and cross-floor exploration without platform-specific learned controllers.
Multi-view target verification: Cross-checks multi-scale observations with a vision-language model, reducing false positives.
Dynamic pedestrian avoidance: Explicitly tracks and predicts moving pedestrians for anticipatory behaviors.
Evaluation: Tested on a real Direct Drive Tech TITA biped robot and extensive simulations on HM3D and MP3D, achieving significant improvements; maintains robust navigation on dynamic benchmark HM3D-DYNA.

More from Embodied