NavHarness: Towards Lifelong Embodied Navigation

NavHarness. Towards Lifelong Embodied Navigation. By Xunyi Zhao and eight others, from Adelaide University, the Responsible AI Research Centre at the Australian Institute for Machine Learning, and one other institution. Published on the twenty-eighth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Frontier models can now do well on a single embodied navigation task, by reasoning over several rounds with multimodal input and a few simple tools. A string of tasks asks for more. The agent has to rely on a map that keeps changing and on records of earlier searches, and either one may be incomplete, or may contradict what the agent now sees. Memory, in other words, is both necessary and unreliable. An agent that ignores it starts from nothing each time, and an agent that trusts it blindly carries old mistakes forward.

NavHarness is a training-free embodied harness aimed at lifelong navigation. Its defining move is to make memory processing part of the navigation loop itself, rather than something kept to one side. Being training-free, it works with the frontier models as they are, and changes what surrounds them rather than what is inside them.

While it navigates, its multi-round agentic session draws on maps, task records and knowledge about the house. It checks them against what it observes and records corrections, and those corrections guide its actions. The experience is then carried across fresh conversations, whether for a new task or for another attempt at a failed one. Outcome verification and summaries written at the end of each run make that experience usable later. The loop, then, has two timescales. Within a session, memory is checked and corrected against what the agent sees. Across sessions, what was learned is verified, summarized and handed on.

On GOAT-Bench, NavHarness improves s-SR over independent sessions that rely on context alone, by eighteen point six points with Astra and by twenty-two point six with Opus 5. Using poses estimated by SLAM, NavHarness with GPT-6 Astra reaches state-of-the-art task success: an s-SR of eighty-three point seven and an e-SR of thirty-six point nine on GOAT-Bench, and an s-SR of eighty-five point nine on IR2R-CE. The first comparison isolates the harness itself: the same models, with and without memory carried between sessions.

To understand where the gains come from, the authors look at how experience is handed from one session to the next. Structured recovery handovers beat summaries matched for length. In extended deployments across houses, consolidation improves navigation beyond simply keeping maps and task records. Case studies show agents using earlier experience to interpret new goals, follow up unresolved questions, and resume searches that had failed.

Better single-task reasoning still counts. The authors' suggestion is that it is only half of lifelong navigation, and the other half lives between tasks: in how each reasoning session builds on what earlier sessions found, got wrong, and wrote down.