- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The goal of the work is to build a harness for lifelong navigation. Lifelong means that, instead of just finding one object in a house and terminating the episode, the agent is tasked with subsequent goals in the same house and can use its experience and memory from prior tasks for the current one. This work approaches the task with multimodal reasoning models and, to better manage memory and context, proposes NavHarness, which orchestrates different module executions and organizes memories from past tasks to best support task execution. In a way, drawing a parallel to coding models, it can be seen as context management for long-horizon tasks. The harness here includes inductive biases specific to indoor navigation. In my impression, NavHarness is an orchestration layer around existing reasoning agents.
Critical strengths
- It clearly improves over naively deploying a multimodal reasoning agent by managing prior memory and context smartly.
- In Table 1, the authors improve over prior work that uses GT poses with SLAM-estimated poses.
- NavHarness gives consistent gains across various model backbones, which suggests that the design is not overfitted to one particular model.
Critical weaknesses
- The paper is hard to follow in places. Although core terms are defined at the start of Sec. 3, the mechanisms in Sec. 3.3-3.4 (recovery scope, the wake-up session, pre-stop verification vs. post-stop certification) are dense and would benefit from a simple worked example. Several details needed to interpret the results appear only in the appendix, e.g., that the reasoning cores run inside Claude Code, Codex, and mini-SWE-agent, and that the baselines receive only observation and action tools (Appendix B.3). Figures 1 and 2 also add little: their captions are brief and generic, and they don’t help explain the mechanisms described in the text.
- Much of the methodology is engineered specifically for lifelong navigation, built around context and memory management on top of existing coding agents. The empirical results are strong, but it is difficult to see how the insights inform other robotics tasks. The paper’s general-sounding findings rely on navigation-specific designs: for example, the structured handover that outperforms a matched-length summary (Table 3) has fields like searched vs. ruled-out places, a reliable landmark, and untried options, and the authors note that this comparison tests information selection and organization together (Appendix B.6). It shows that this hand-designed structure helps, rather than a general principle about what should carry across sessions.
- The memory practices themselves (persistent files, handovers, compaction, agent-directed retrieval) are adopted from coding and memory agents such as MemGPT, as the paper acknowledges. The harder and more general problem is how to learn what experience to retain and how to compress it, which this work specifies by hand.
- Many parts of the pipeline are hand-defined based on task-specific observations, so it is unclear whether there is a general design scheme to follow or whether the system reflects navigation-specific insights gathered from extensive experimentation.
- Is some of the handoff and compression just prompt engineering?
- The independent-session baselines run in Claude Code and Codex but receive only observation and action tools (Appendix B.3). Would a generic coding agent given the same map and file tools and a persistent workspace, instructed to keep its own notes, recover much of the gain without NavHarness’s orchestration? This would show how much of the improvement comes from the harness design versus simply having persistent tools, and whether improving coding agents might make task-specific harnesses unnecessary.
Questions for authors
- The baselines share NavHarness’s per-task budgets of 200 navigation-model turns and 500 steps (Table 11), but NavHarness also makes model calls outside this budget for the judge, consolidation, and recovery wake-up session, and tasks use only around 40 turns on average, so the turn cap rarely binds. Could the authors report total model calls and tokens per task for NavHarness and the baselines? This would show whether part of the gain comes from additional inference compute.
- I would be very interested to hear how this specific harness provides insights on how to improve context and memory management for robots. It doesn’t need to be experiments, but even an interesting insight would help. Without it, I lean towards a rejection.
A kind follow-up on our response
Dear Reviewer aCkT,
We sincerely thank the reviewer again for the time and care devoted to our paper. As the discussion period is coming to an end, we would like to kindly ask whether our response has addressed the reviewer's concerns. We have revised the paper accordingly, with all changes marked in blue and summarized in the general response. If any points remain unclear, we would be glad to provide further clarification before the discussion closes.
We are very grateful for the reviewer's thoughtful feedback.
Best regards, The Authors