acceptodds
Withdrawn from ICLR 2027

EONav: Support-Centric Memory and Active Search for Everyday Object Agentic Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Everyday object navigation spans space and time: small targets may require close inspection of plausible supporting surfaces, while portable objects can relocate between searches. An agent must therefore decide where to inspect, how to use potentially stale sightings, and which new observations to trust. We introduce an Everyday Object Navigation benchmark with 560 static episodes and 600 ordered dynamic episodes across 35 scenes, jointly evaluating fine-grained discovery and repeated search under relocation. We further propose EONav, a training-free, VLM-centered framework that connects memory, planning, and active verification through relatively stable supports. Its memory agent retains multiple verified, timestamped target–support observations; their historical frequencies provide an empirical support-level prior, with recency as an additional cue. A VLM planning harness combines this history with current observations and inspection feedback to select supports to revisit, falling back to support-guided exploration when history is unavailable or exhausted. Active verification obtains closer or alternative views, redirecting search after rejections and admitting only confirmed observations to memory. EONav achieves 54.46% success rate (SR) and 26.58% success weighted by path length (SPL) in static search, and 61.50% SR and 40.14% SPL in dynamic search, outperforming the evaluated baselines. Temporal memory raises dynamic SR from 23.26% to 61.50%, while EONav uses 23.1% fewer tokens per episode than STRIVE. We also demonstrate real-robot object search and will release the code and benchmark.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Agentic Embodied Navigation, Everyday Object Navigation, Spatial-Temporal Memory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sVQAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce an Everyday Object Navigation benchmark which includes a static set and a dynamic set. They also propose EONav, a training-free VLM agent built on STRIVE that adds three key features: support-anchored temporal memory to track target-support sightings, a VLM-guided planning agent for revisiting supports or prioritizing relevant frontiers, and active verification before stopping or committing detections to memory.

Critical strengths
  • Reproducible Benchmark Construction and Realistic Problem Framing. The paper targets a key gap in everyday object navigation by benchmarking three combined challenges: revisiting relocated objects, inspecting tight spaces for small targets, and managing long-term memory updates. The dataset construction is rigorous and reproducible.

  • Validating Experiments with Performance Gain. Cross-episode temporal memory drives the dynamic search performance, more than doubling success rates while holding the exploration and verification stack constant. The results support maintaining multi-record histories rather than overwriting past sightings, which helps recover targets that return to earlier locations.

Critical weaknesses

The ablations don’t explain where the gains come from

  • It is difficult to tell what actually drives EONav’s performance gains over STRIVE. Even with all support-guided features turned off, EONav still beats STRIVE by over 11 points in static search (52.14% vs. 40.71%), suggesting the gap may stem from using a stronger model stack (Gemini 3.7 Flash, GroundingDINO, SAM) rather than the proposed method itself.
  • Key components are either unablated (active verification) or provide negligible gains (the frequency/recency prior), while basic control baselines are missing, such as simply revisiting the nearest known support first. To validate the claims, I would suggest to evaluate STRIVE on the same model stack, active verification needs a isolated ablation, and statistical noise should be tested, as several reported 1–3% ablation differences seem to fall within normal single-run VLM variance.
Questions for authors
  • Which VLM and detector does STRIVE use in your runs, and what does STRIVE score with gemini-3.7-flash?
  • What is SR/SPL with active verification removed?
  • What exactly does the w/o-memory ablation remove, and why does the no-history subgroup score 47.65 when the ablation scores 23.26?
  • How are pre-existing scene instances of the target category handled at evaluation?
  • Fig. 11 (ep. 0403) shows the agent stopping at a pre-existing bowl-like scene object and being scored as a failure. Do HM3D scenes contain native instances of the target categories? If so, are those excluded at placement time, or counted as valid goals? Otherwise category-level successes may be scored as failures.
Reviewer_yTNpOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies everyday-object navigation, combining fine-grained object discovery with repeated search after inter-episode relocation. It introduces a benchmark comprising 560 static episodes and 600 dynamic episodes organized into 50 sequences across 35 HM3D scenes. The proposed training-free framework, EONav, maintains verified target-support histories, aligns historical supports with current observations, and combines memory-guided revisitation with support-guided exploration. Active verification informs subsequent search and controls memory updates. The reported results show improvements over the evaluated baselines in both success rate and SPL, together with reduced token usage relative to STRIVE. Qualitative real-robot demonstrations illustrate deployment for everyday-object search.

Critical strengths
  1. Substantial end-to-end improvements on a practically motivated benchmark. The proposed benchmark combines fine-grained object discovery with repeated search after relocation, addressing a relevant capability for household robots. EONav achieves substantial improvements over the evaluated baselines: static SR increases from STRIVE’s 40.71% to 54.46%, and dynamic SR from 17.58% to 61.50%, with accompanying improvements in SPL. These results provide promising evidence that the integrated pipeline improves both search success and navigation efficiency on the proposed task.
Critical weaknesses
  1. The experiments do not sufficiently establish the distinctive value of the proposed memory and verification designs. Existing work already explores several relevant alternatives, including anchor-based associations in DualMap, dynamic memory updates in DynaMem, and long-term memory coupled with action feedback in STAR. Although DualMap is included as a system-level baseline, the last two rows of Table 2 primarily establish the benefit of cross-episode memory, rather than the advantage of the proposed support-centered, multi-observation representation over simpler or established alternatives. The Recency baseline in Table 3(a) retains all historical candidates and only changes their order; it therefore does not replace a latest-sighting-only memory control. The contribution of verification-guided search and memory updates is similarly underexamined. Removing local inspection in Table 3(b) still retains target confirmation, and the paper does not directly measure whether the proposed verification procedure reduces incorrect stopping decisions, erroneous memory writes, or their effects on subsequent searches. Moreover, disabling all three support-search components still yields 52.14% static SR, compared with 40.71% for STRIVE and 54.46% for the full system, leaving most of the headline SR improvement unattributed. Matched comparisons that isolate historical retention, support anchoring, and active verification would provide stronger evidence for the claimed methodological contribution.

  2. The real-robot demonstrations are insufficiently documented to assess the transfer of the central memory mechanisms. Section 5.4 and Figure 7 illustrate searches for a pen, remote control, and keyboard, but do not establish whether these trials exercise cross-episode memory retention, object relocation, or historical-support alignment. The provided appendix does not supply additional real-robot evidence resolving these questions. Consequently, the demonstrations establish deployment feasibility more clearly than they validate the proposed memory mechanism under physical-world conditions. Even within a qualitative evaluation, the paper should describe the memory initialization and persistence, localization and sensing setup, and any deployment-specific modifications, together with representative successes and failures. A focused discussion should explain whether the system observed real-world deployment issues, how it responded, and which remain potential limitations. This would clarify the conditions under which the memory and verification design transfers beyond simulation.

Questions for authors

If possible, please add corresponding ablations to highlight the quantitative gain of the proposed technical contribution. Please also provide more details and insights from your real-world experiments.

Reviewer_RxwPOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a benchmark for searching for everyday objects, including repeated searches after objects relocate, and proposes EONav, a training-free agent that uses support-centered temporal memory, support-guided exploration, and active verification. It reports 54.46% SR and 26.58% SPL on static search, and 61.50% SR and 40.14% SPL on dynamic search. Removing temporal memory reduces dynamic SR to 23.26%.

Critical strengths
  1. The paper connects its proposed memory and planning mechanisms to a benchmark that tests both small-object discovery and repeated search after relocation. This makes the contribution useful as both a method and an evaluation setting.
  2. The reported gains are supported by component ablations. In particular, removing temporal memory substantially reduces dynamic-search performance, supporting the paper’s central claim about memory reuse.
Critical weaknesses
  1. The 1.0 m success distance is coarse for a paper centered on fine-grained search for small objects. The reported SR and SPL therefore do not establish how reliably the agent reaches or identifies the object at close range.
  2. Dynamic episodes are organized into 50 sequences, and the paper does not report sequence-level uncertainty or confidence intervals. The static benchmark’s category counts are also highly uneven, which may affect how representative the aggregate results are.
Questions for authors
  1. Do the main static and dynamic gains persist under a stricter success distance? If such results are available, what are the SR and SPL at that distance?
  2. What are the sequence-level results or confidence intervals for dynamic search, accounting for the 12 related episodes within each sequence?
  3. Could the authors provide a small sample of the data, as well as a timeline for the open-source release of the benchmark?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.