NavHarness: training-free harness for lifelong embodied navigation, 83.7 s-SR on GOAT-Bench
Xunyi Zhao · hf · 2026-10-01
NavHarness: lifelong embodied navigation
- Frontier models handle single navigation tasks, but successive tasks require an evolving map and search records that may be incomplete or conflict with new observations.
- NavHarness is a training-free embodied harness making memory processing part of the navigation loop: multi-round agentic sessions check maps, task records and house knowledge against observations, record corrections, and preserve experience across fresh conversations for new tasks or recovery, with outcome verification and run-end summaries.
- Gains: +18.6 s-SR (Astra) / +22.6 (Opus 5) over context-only sessions on GOAT-Bench; GPT-6 Astra with SLAM poses achieves SOTA 83.7 s-SR / 36.9 e-SR, and 85.9 s-SR on IR2R-CE.
- Structured recovery handovers beat length-matched summaries; consolidation improves navigation beyond just retaining maps and records.
More from coding & agent
- Agent autonomously checks in a flight 24 hours ahead — and pings only on failure — armand_ruiz · 2026-10-01
- 0.8B model plus 9 LoRA adapters routes agent decisions 38x faster with +8.7 accuracy — Usual_Maximum7673 · 2026-10-01
- Context engineering's next step: knowledge system engineering — ShanRizvi · 2026-10-01
- Codex's model picker now needs a scrollbar — should OpenAI add auto-routing? — darkprinceimmortal · 2026-10-01
- OpenAI's DevDay agent launches put startups on notice, says founder in the crosshairs — NoSpecific64 · 2026-10-01
- Dev adds Apple's Genie effect to AI agent plugin with Opus, says he's basically built an agentOS — RileyRalmuto · 2026-10-01