HTCE: Homeomorphism-inspired Topological Cross-Scene Experience Canonicalization and Reuse for Object Navigation
| Source: arXiv AI
Tags: HTCE, Object-Goal Navigation, HM3D, MP3D, Embodied Agents, Vision-Language Models
HTCE lets embodied navigation agents reuse past successful routes across different homes by canonicalizing their topology, topping SR and SPL on HM3D and MP3D with a frozen vision-language policy.
Details
Robots doing object-goal navigation build up successful experiences whose semantic structure repeats across homes even when the paths differ. Naming and granularity differences make similar approaches look different as graphs. HTCE, inspired by homeomorphism, applies feature and topology subdivisions and then reduces them into comparable semantic sequences and structural signatures, keeping anchors, branches and endpoints intact. Matching retrieved past routes then guides a frozen vision-language policy. Canonicalization merges 17.7–23.6% of stored route descriptions without losing task performance. With the same frozen backbone for all systems, HTCE reaches 59.9%/31.8% SR/SPL on HM3D v0.1, 71.0%/31.5% on HM3D v0.2 and 46.6%/16.6% on MP3D, with SR gains up to 4.9 points. Code is available.