AI
NavHarness: Towards Lifelong Embodied Navigation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 14:35:41 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 13:18:23 UTC.
Briefings
NavHarness is a training-free embodied navigation framework that enables lifelong navigation by integrating multi-round agentic sessions with evolving maps and task records. On GOAT-Bench, it achieves state-of-the-art performance on s-SR (83.7) and e-SR (36.9) using SLAM-estimated poses and GPT-6 Astra, surpassing context-only sessions by 18.6 points. The system preserves experience across sessions through structured recovery handovers, which outperform length-matched summaries. Results from extended house deployments demonstrate that consolidation of prior experience allows agents to interpret new goals and resume failed searches, suggesting progress depends on both session building and single-task capability.
Key points
- NavHarness uses multi-round agentic sessions with maps, task records, and house knowledge to guide actions during navigation.
- On GOAT-Bench, NavHarness with GPT-6 Astra achieves 83.7 s-SR and 36.9 e-SR, outperforming context-only sessions by 18.6 points.
Context and limitations
Results are limited to GOAT-Bench and IR2R-CE benchmarks; extended house deployments show promise but case studies are limited to specific agents and tasks.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 14:35:41 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiaxhxjyb7xbtokc4ocknmunfrqtioiclahtjgcfkhmie7ewiel644