PaperlayerView exact revision on arXiv

Source-linked AI summary

NavHarness: Towards Lifelong Embodied Navigation

Xunyi Zhao, Jian Zhou, Sihao Lin, Gengze Zhou, Zerui Li, Xinyu Yan, Jiajun Liu, Anton van den Hengel, Qi Wu

arXiv:2609.34276v1cs.RO

TL;DR

Success on individual navigation tasks does not resolve how an agent should use evolving maps and incomplete or conflicting search records across successive tasks. NavHarness integrates memory processing into multi-round navigation sessions and carries experience across fresh conversations; it improves benchmark performance and structured recovery handovers outperform length-matched summaries.

  • Problem

    Across successive navigation tasks, agents must use evolving maps and earlier search records that may be incomplete or conflict with new observations.

    Source: §1 · p. 1

  • Method

    NavHarness uses multi-round sessions to consult observations, spatial and task memory, and house knowledge, while preserving records across fresh task or recovery conversations.

    Source: §1 · p. 2

  • Results

    On GOAT-Bench, NavHarness improves s-SR over context-only independent sessions by 18.6 points with Astra and 22.6 with Opus 5.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Structured recovery handovers outperform length-matched summaries, and extended deployments show consolidation improves navigation beyond retaining maps and task records.

    Source: §1 · pp. 2–3

    Continuous-deployment evidence covers finite task sequences in multiple static simulated houses; extending evaluation to changing environments and longer deployments remains important.

    Source: E Limitations and Future Directions · p. 51

Abstract

from arXiv · show

Frontier models can now perform well on individual embodied navigation tasks through multi-round multimodal reasoning with simple tools. Across successive tasks, however, an agent must also rely on an evolving map and earlier search records, both of which may be incomplete or conflict with new observations. We present NavHarness, a training-free embodied harness towards lifelong navigation that makes memory processing part of the navigation loop. During navigation, its multi-round agentic session draws on maps, task records, and house knowledge, checking them against observations and recording corrections to guide its actions. NavHarness preserves this experience across fresh conversations for new tasks or recovery attempts, while outcome verification and run-end summaries support its later reuse. On GOAT-Bench, NavHarness improves s-SR over context-only independent sessions by 18.6 points with Astra and 22.6 with Opus 5. Using SLAM-estimated poses, NavHarness with GPT-6 Astra achieves state-of-the-art task success of 83.7 s-SR with 36.9 e-SR on GOAT-Bench and 85.9 s-SR on IR2R-CE. To understand these gains, we examine how experience is carried between sessions and find that structured recovery handovers outperform length-matched summaries. In extended deployments across houses, consolidation improves navigation beyond retaining maps and task records, with case studies showing how agents use earlier experience to interpret new goals, investigate unresolved questions, and resume failed searches. We suggest that progress towards lifelong navigation depends on how successive reasoning sessions build on prior experience, alongside improvements in single-task capability.

1 Introduction

Lifelong navigation requires agents to use and revise accumulated maps and search records against new observations, not merely retain them. NavHarness puts this memory processing into navigation sessions and preserves experience across tasks and recovery attempts.

1 Introduction

  • A task record that an object was not found does not establish that every relevant part of a room was inspected, and stored descriptions may conflict with current views.The navigator must interpret records against observations and revise them when evidence changes.Source: §1 · p. 1
  • NavHarness combines current observations, spatial and task memory, and house knowledge in multi-round sessions, while preserving records across fresh conversations without additional model training.Outcome checks and run-end consolidation support later reuse of the records.Source: §1 · p. 2
  • On GOAT-Bench, NavHarness improves s-SR over context-only independent sessions by 18.6 points with Astra and 22.6 with Opus 5.With SLAM-estimated poses, GPT-6 Astra reaches 83.7% s-SR and 36.9% e-SR on GOAT-Bench, and 85.9% s-SR on IR2R-CE.Source: §1 · p. 2
  • Structured recovery handovers outperform length-matched summaries, bringing evidence transfer into a navigation loop where fresh sessions continue from the robot’s current state.The handovers carry evidence from earlier searches.Source: §1 · p. 2
  • Continuous deployment shows consolidation improves navigation beyond retaining maps and task records, with case studies of agents reusing failed searches and revising earlier accounts.The authors connect these findings to the design of future long-horizon embodied agents.Source: §1 · p. 3

2 Related Work

Prior work addresses long-horizon navigation through persistent scene knowledge, retrieval, planning, and continued learning. NavHarness focuses on how general-purpose multi-round sessions consult and revise several kinds of memory across session boundaries.

Long-horizon and lifelong navigation

  • Long-horizon navigation work retains scene knowledge, retrieves past observations, or learns policies that query accumulated experience; NavHarness studies general-purpose sessions using these sources together.Examples include structured objects, visual snapshots, semantic graphs, renderable memories, hierarchical planning, active retrieval, and continued policy learning.Source: §2 · p. 3

Embodied harnesses and agentic control

  • Embodied harnesses connect multimodal reasoning to navigation tools, robot controllers, or reusable skills, while NavHarness studies what evidence must cross a reasoning-session boundary.That evidence includes inspected places, what observations exclude, and where search remains incomplete.Source: §2 · p. 3

Memory processing in long-horizon agents

  • Long-horizon agents use external memory, revisable records, compression, and tool-mediated file access; NavHarness tests agent-directed access and structured search handovers for physical search.Checking records in physical search may require revisits.Source: §2 · p. 3

3 NavHarness

NavHarness organizes navigation as task and recovery sessions that use persistent spatial, task, and house memory, with an orchestrator managing transitions. Verification and run-end consolidation prepare experience for reuse across tasks and runs.

3 NavHarness

  • The orchestrator manages sessions, recovery, completion requests, and retained memory, advancing sessions in bounded hops while tracking task budgets and robot state.A run is a sequence of tasks in one house; recovery replaces the conversation while retaining physical state, memory, and remaining budget.Source: §3 · pp. 3–4

3.1 Starting a task with retained experience

  • Fresh task conversations receive references to prior handovers, the run ledger, and house notes, which they read through file tools before moving.Working memory pairs task records with SLAM-based spatial memory; house notes cover routes, failed searches, evidence-supported exclusions, and unresolved questions.Source: §3.1 · p. 4

3.2 Navigating by consulting and updating memory

  • Sessions compare retrieved house notes, task records, and map queries with new observations to decide whether to follow an earlier account or seek more information.The growing occupancy map supports inspection of explored space and route previews, with named places linked to saved views.Source: §3.2 · p. 5

3.3 Recovering with a fresh conversation

  • Recovery replaces an unproductive attempt with a fresh conversation that reads a note of searched places, evidence-backed exclusions, untried options, and uncertainty.The new attempt retains the map, other records, location, and remaining budget; a visited room is not assumed to have been fully searched.Source: §3.3 · p. 5

3.4 Checking completion and passing on the task record

  • A separate judge checks completion using goal and visual evidence, and a contradictory definite pre-stop verdict can return the task to search before STOP.Post-stop certification records the final assessment without changing the score; claims and verdicts remain separate because a check does not validate every spatial assertion in a handover.Source: §3.4 · p. 6

3.5 Consolidating experience for the next run

  • At run end, a consolidation session rewrites four house-note files from the journal and existing notes, integrating supported experience while preserving source records.It cannot change map geometry or overwrite task accounts; later runs read the notes alongside saved maps and views.Source: §3.5 · p. 6

4 Experiments

The experiments evaluate NavHarness on successive-goal benchmarks and compare it with independent sessions and prior methods. Results report higher task and sequence success, while the component study is designed to compare session policies, memory, recovery, and completion checks.

4.1 Experimental setting

  • The evaluation uses validation-unseen GOAT-Bench and IR2R-CE splits, with memory carried between subtasks within an episode or tasks within a tour, but reset between episodes or tours.GOAT-Bench includes 2,669 subtasks in 360 episodes across 36 scenes; four reasoning cores receive the same navigation and memory tools.Source: §4.1 · p. 6

4.2 Comparison with state-of-the-art methods

  • NavHarness improves GOAT-Bench s-SR over independent sessions by 30.3 points for Qwen3.8-27B, 34.8 for GPT-4o, 22.6 for Opus 5, and 18.6 for GPT-6 Astra.For Astra, e-SR rises from 13.6% to 36.9% on GOAT-Bench and from 9.3% to 27.8% on IR2R-CE.Source: §4.2 · pp. 6–7
  • With SLAM-estimated poses, GPT-6 Astra reaches 83.7% s-SR on GOAT-Bench, 62.3 SPL there, and 85.9% s-SR on IR2R-CE, exceeding prior scores cited for each benchmark.The table captions identify the GOAT-Bench and IR2R-CE comparisons; the prose gives the values and comparisons.Source: Table 1 · p. 7, Table 2 · p. 7, §4.2 · p. 7

4.3 What carries navigation across tasks and attempts?

  • The Opus 5 component study compares session policies, cross-task memory, recovery handovers, and completion checks on the full GOAT-Bench validation-unseen split.Recovery shares the original task’s budget.Source: §4.3 · p. 8, Table 3 · p. 8

A. Session policies

The two supplied independent-session rows report distinct numeric sequences; the passages do not identify the metric labels for those values.

A. Session policies

  • Independent sessions are reported as 58.9 ± 0.9, 34.9 ± 0.8, and −22.6 [−26.3, −19.0].Source: A. Session policies · p. 8
  • The first reported value is 3.5 higher with pre-stop verification than without it (62.4 versus 58.9).Source: A. Session policies · p. 8
  • Independent sessions with pre-stop verification are reported as 62.4 ± 0.8, 36.5 ± 0.7, and −19.1 [−22.4, −16.4].Source: A. Session policies · p. 8

B. Cross-task memory

The supplied rows give numeric results for three cross-task memory conditions, without identifying the metric labels for each column.

B. Cross-task memory

  • Clearing the map while keeping records is reported as 68.9 ± 0.8, 39.8 ± 0.7, and −12.6 [−15.6, −9.7].Source: B. Cross-task memory · p. 8
  • Hiding earlier task records while keeping the map is reported as 71.8 ± 0.7, 41.8 ± 0.6, and −9.7 [−12.5, −7.0].Source: B. Cross-task memory · p. 8
  • No cross-task memory is reported as 66.8 ± 0.9, 37.4 ± 0.7, and −14.7 [−16.9, −12.4].Source: B. Cross-task memory · p. 8

C. Recovery handovers

The supplied rows report results for four recovery conditions, but do not identify the metric labels for each column.

C. Recovery handovers

D. Completion checks

The section reports that consolidation improves continuous-deployment results beyond retained maps and task records, while examining how navigation sessions reuse and revise experience. It also identifies remaining risks in memory reliability.

D. Completion checks

  • Hiding task records costs 9.7 s-SR points and increases steps by 25%, while clearing spatial memory costs 12.6 points and increases steps by 32%.Without cross-task memory, s-SR falls by 14.7 points; later tasks must reconstruct routes or repeat searches.Source: D. Completion checks · p. 8
  • A matched-length ordinary recovery summary trails the structured handover by 8.3 points; the handover preserves evidence about ruled-out and untried search locations.Source: D. Completion checks · p. 8
  • Completion checks have smaller s-SR effects than memory used throughout a search, while post-stop certification raises record accuracy to 90.7% from 82.6%.Pre-stop verification can prolong the search; certification changes the record inherited by later tasks.Source: D. Completion checks · pp. 8–9

4.4 Continuous deployment

  • The continuous-deployment evaluation joins ten tours per house across simulated days, placing the robot at each next start to model resumption after shutdown.Source: §4.4 · p. 9
  • Across 36 houses, consolidation raises pooled s-SR from 72.8% to 80.5% and SPL from 36.5 to 44.3, with paired gains of 7.7 and 7.8 points.Source: §4.4 · p. 9, Figure 3 · p. 9
  • House notes connect ambiguous goal wording to intended objects, and sessions can investigate unresolved memory questions to inform later interpretations.One session revisits whether two refrigerator descriptions refer to one or two appliances, then records its findings for consolidation.Source: §4.4 · p. 9
  • Detailed failed-search records can help a later session resume at an unchecked location, while new observations can correct only the advice attached to retained spatial knowledge.Source: §4.4 · pp. 9–10
  • A judge-accepted incorrect completion can propagate into house notes, and correcting one record does not necessarily update copies or validate other spatial claims.Source: §4.4 · p. 10

5 Conclusion

NavHarness carries experience across fresh navigation sessions and revises it against observations, supporting navigation across multiple reasoning models. Its harness coordinates session handovers, memory, tools, and outcome checks across tasks.

5 Conclusion

  • NavHarness lets multi-round sessions use maps, search records, and house knowledge without additional training, carrying experience forward while checking and correcting it against new observations.The conclusion highlights uses of prior experience to interpret ambiguous goals, investigate unresolved questions, and resume failed searches.Source: §5 · p. 10

A Harness Specification

  • Across a sequence of tasks, the orchestrator mounts memory, opens fresh sessions, verifies closure claims, records outcomes, and can consolidate house notes at run end.The algorithm also supports recovery by retaining the goal, physical state, map, task records, and remaining budget while replacing the conversation.Source: A Harness Specification · p. 15

A.1 The loop

  • A coding-agent session interprets observations, consults memory, and invokes tools; the orchestrator manages task progression and whether the session continues, closes, or hands experience to a fresh attempt.The boundary state includes session context, occupancy grids, named places and views, a place graph, task records, and persistent house notes.Source: §A.1 · p. 16

A.2 Tool registry

  • The tool registry gives the navigator movement, mapping, and route queries, while the harness adds file access and closure and recovery requests; maps and named place markers persist across tasks.The spatial tools provide map views, place markers with associated views, and route previews without moving the body.Source: §A.2 · pp. 16–17
  • The occupancy grid is built from depth and online pose estimates, so later tracker loop-closure corrections are not propagated into cells already written.The map uses 0.10 m cells and distinguishes unknown, free, and obstacle states.Source: §A.2 · p. 18

A.3 Context management

NavHarness separates task and recovery sessions from persistent memory, letting fresh conversations reuse maps, task records, and house knowledge. Recovery scopes preserve accumulated state, while verification and consolidation govern what later sessions inherit.

A.3 Context management

  • NavHarness opens fresh conversations at task boundaries while retaining external memory; task-level contexts average approximately 20k tokens, while single session reaches a roughly 1M-token window near task 20.The backend compacts the single-session conversation and continues; this boundary does not end the deployment. Under the stated input-price assumptions, 20-task input costs are approximately $2,520 for stateless rebuilding and $98 for NavHarness.Source: §A.3 · pp. 19–20

A.4 Memory

  • Working memory holds maps, places, and task records, while long-term house memory adds notes and experience archives; the same map can serve both lifetimes without textual duplication.Task handovers record goals, visited places, outcome evidence, and unsuccessful searches; the ledger indexes closed tasks, while detailed records remain separately accessible.Source: §A.4 · pp. 20–21
  • At run closure, consolidation summarizes task records into house notes for later runs, but cannot modify map geometry and does not certify stored claims.The consolidation session reads the ledger, handovers, and previous notes, and can retain routes, failed searches, corrections, and unresolved questions.Source: §A.4 · p. 22, §A.5 · p. 24

A.5 Hooks

  • Recovery replaces the conversation while preserving physical trajectory, remaining budget, map, markers, ledger, and archive; broader scopes reconsider more beliefs rather than resetting state.The controller uses measurable execution signals, including new explored area and an observation-count proxy, rather than an oracle for the cause of failure.Source: §A.5 · p. 23
  • Recovery handovers distinguish searched places from evidence-supported exclusions, but those exclusions remain the actor’s interpretations and a mistaken one can be carried into the next attempt.For scopes above attempt, a separate wake-up session can compare stored place photographs with current views and provide a location brief.Source: §A.5 · pp. 23–24

A.6 Verification

  • Pre-stop verification can return counter-evidence to the actor for another search, while post-stop certification assigns the stored task status from final evidence.The one-time pre-stop check cannot withhold closure indefinitely; if action-step count is unchanged, certification reuses the pre-stop verdict.Source: §A.6 · pp. 24–25
  • Actor and judge errors are not independent merely because their model parameters match; the checks instead differ in evidence access and in which records the actor may write.Environment scores enter evaluation logs, not the model-based judge or the agent-facing ledger.Source: §A.6 · p. 25

B Benchmarks and Evaluation Protocol

The evaluation covers sequential navigation on GOAT-Bench and IR2R-CE, with controlled comparisons, paired statistics, and interventions on memory, recovery, and verification. Results show gains over independent sessions and benefits from agent-directed retrieval, while comparison protocols and evaluation scope shape what the results establish.

B.1 GOAT-Bench; B.2 IR2R-CE

  • GOAT-Bench evaluates sequential multimodal navigation across 2,669 subtasks, 360 episodes, and 36 scenes; IR2R-CE covers 1,824 tasks across 36 tours and 11 scenes.Each benchmark run starts with empty memory, which persists between tasks but is not inherited by another run; the navigator uses visual-SLAM pose rather than simulator localization.Source: B Benchmarks and Evaluation Protocol · p. 25, §B.1 · p. 26, §B.2 · p. 26

B.3 Comparison sources and settings

  • The four reasoning executors share NavHarness’s mapping, memory, and recovery mechanisms, while same-backbone controls compare surrounding harness policies rather than isolating memory alone.Published baselines may use different perception and control pipelines or evaluation protocols, so matching model names does not establish equivalent systems.Source: §B.3 · pp. 26–27, §B.6 · p. 29

B.4 Statistics; B.5 Chain success

  • Three-seed comparisons report means and standard deviations; paired component intervals use scene-clustered resampling, while chain success requires every task in an episode or tour to succeed.Component intervals are pointwise and lack multiple-comparison adjustment.Source: §B.4 · p. 28, §B.5 · p. 28

B.6 Interventions

  • The controlled interventions independently vary complete session policies, cross-task memory, recovery and handover content, and verification functions or stop-view evidence.The cross-task memory controls distinguish persistent maps from earlier task records; recovery-note comparisons retain recovery triggers, scope control, and shared budgets.Source: §B.6 · pp. 28–29

B.7 Memory access and judge choice

  • Agent-directed historical access raises s-SR by 7.0 percentage points over fixed access (95% CI [+5.9, +8.2]), with SPL increasing from 49.0 to 55.0.The comparison tests query timing, selection, and follow-up together; fixed access retrieves original excerpts on a schedule within a 2,000-token budget.Source: §B.7 · p. 29, Table 12 · p. 30
  • Swapping the completion-check judge changes s-SR by −0.4 points for Qwen and +0.2 for Opus, leaving navigation performance close to default in both configurations.Similar navigation scores do not establish equal judgment accuracy or exclude errors shared by the two models.Source: §B.7 · p. 30, Table 13 · p. 30

C.1 Component effect profile

  • Figure 9 compares controlled Opus interventions against the full harness, with paired uncertainty; every interval shown excludes zero and positive values favor the full harness.The component comparisons use the same full-split task set, and each intervention starts independently from the reference configuration.Source: §B.6 · pp. 28–29, §C.1 · p. 30, Figure 9 · p. 31

C.2 Session and system comparisons

  • With GPT-6 Astra held fixed, NavHarness improves s-SR by 18.6 points, SPL by 22.2 points, and e-SR by 23.3 points over independent sessions; Qwen3.8-27B and GPT-4o also improve on all three measures.In the full-split Opus study, independent sessions lose 22.6 success points relative to full, while a single session loses 26.6 points.Source: §C.2 · p. 31

C.3 Within-task recovery and handovers

Structured recovery notes outperform length-matched ordinary summaries, and cross-task memory contributes to navigation success. Recovery-event success rates are conditional on the events observed in each intervention, whereas the intervention comparison tests overall task success.

C.3 Within-task recovery and handovers

  • The matched-length ordinary-summary control reaches 73.2% s-SR and 47.0% SPL, compared with 81.5% and 55.0% for the full harness; its s-SR change is −8.3 percentage points (95% CI [−11.0, −5.7]).This comparison tests structured recovery notes against general summaries, alongside empty-note and task-mismatched-note interventions.Source: §C.3 · p. 31
  • Among recorded recovery events, 56.1% lead to success with the attempt’s own note, versus 31.7% with an empty note and 35.0% with another task’s note.The model requests recovery at a median of 32 turns versus 84 for the fallback; request-triggered recoveries succeed 61.5% of the time versus 52.5% for fallback-triggered recoveries, though the triggers select different failures.Source: §C.3 · pp. 31–32, Figure 10 · p. 32

C.4 Cross-task memory

  • Disabling cross-task memory while retaining within-task mapping, recovery, and verification yields 66.8% s-SR and 37.4% SPL; the paired s-SR change is −14.7 points (95% CI [−16.9, −12.4]).Without the persistent map, mean steps rise to 168 from 127 with the full harness, markers per task rise to 3.4 from 1.3, and recovery is requested 1.8 times as often.Source: §C.4 · p. 32

C.5 Verification uses and evidence

Pre-stop verification contributes to navigation success, while post-stop certification distinguishes supported outcomes from unsupported completion claims. The diagnostics also define limits on what can be inferred from the verification comparisons.

C.5 Verification uses and evidence

  • Disabling pre-stop verification reduces full-system success from 81.5% to 76.1%, a 5.4-point change, and SPL from 55.0% to 47.0%.With verification enabled in both configurations, the full system exceeds independent sessions by 19.1 success points; the remaining harness features are evaluated jointly.Source: §C.5 · p. 32, Table 15 · p. 33
  • The local diagnostic estimates the current-task consequence of continuing after a refused first STOP, conditional on histories produced with pre-stop verification enabled.It does not estimate what would happen if an entire task sequence were rerun without pre-stop verification.Source: §C.5 · p. 32
  • After false records, successor-task success is 66.7% with certification versus 61.4% without; after true records, it is 82.7% versus 81.6%.These rates condition on each configuration’s histories and do not isolate certification’s effect.Source: §C.5 · pp. 32–33, Table 16 · p. 33
  • Four-view accuracy is 90.7% versus 82.6% for accept-all, while single-view accuracy is 82.9% versus 80.6%.The four-view checker rejects 70.1% of false claims and accepts 95.0% of true claims, but 25.2% of its rejections concern true completions.Source: §C.5 · p. 33

D.1 Component effects and failure diagnostics

  • Component-effect diagnostics help connect aggregate intervention losses to observed failure patterns, but do not estimate effects within difficulty or failure-type strata.The diagnostics are not separate effect estimates for tasks grouped by those conditions.Source: §D.1 · p. 33

D.2 With and without consolidation across multiple houses

The extended deployments compare consolidation with retaining maps and task records alone. Consolidation’s observed gains vary by house and goal type, and the case-study trajectories are not decomposed into independent effects of individual records or tools.

D.2 With and without consolidation across multiple houses

  • The deployment schedule extends navigation to ten tours across simulated days, with tour-boundary relocations modeling work resumed after shutdown.The full configuration adds run-end house-note consolidation; both configurations retain maps and task records, and the detailed comparison covers three houses from a 36-house evaluation.Source: §D.2 · p. 34
  • For described goals, success rises from 77% to 95% in House A and 60% to 72% in House C, while steps per task fall from 217 to 93 and 204 to 157.House B’s object goals are all completed by both runs, with 66 versus 41 steps per task.Source: §D.2 · p. 35
  • House A improves in success and route efficiency; House B’s tasks closed within ten steps rise from 8% to 18%; House C gains almost entirely in success.For House C, the apparent lack of efficiency gain on tasks both runs complete reflects chained starts; efficiency rises when starts coincide.Source: §D.2 · pp. 34–35, Figure 13 · p. 36
  • Keyword matches in task accounts do not establish that a memory was read or revised; both configurations retain markers.The figure also counts re-placing an existing marker under an existing label as an action, not a phrase.Source: Figure 14 · p. 36
  • Task-level trends differ by house: House A’s full-run success dips below control only on day two, House B’s larger difference is in steps, and House C gains on tasks flipped to success.The fourteen cases were selected from forty-eight analyzed cases to illustrate behaviors, not estimate their frequency across navigation tasks; trajectory differences are not decomposed by tool or record.Source: §D.2 · p. 35, Figure 15 · p. 37, Figure 16 · p. 37

D.2.1 Revising spatial records (Cases 1 to 4)

The spatial-record cases show that re-placing markers can correct coordinates after frame changes, but inherited labels and routes can still mislead. Repeated visits also refine marker locations, without guaranteeing shorter searches on every task.

D.2.1 Revising spatial records (Cases 1 to 4)

  • In House A, 13 of 29 marker re-placements were frame events rather than moves confirmed by the simulator; Houses B and C show the same pattern.In House A, nine frame events occurred at day starts and four after within-day tracking loss; seven corrections were themselves placed in a broken frame. Houses B and C had 11 of 15 and 21 of 34 frame events, respectively.Source: §D.2.1 · p. 38, Figure 17 · p. 38
  • After a day-start reset displaced the laundry-bath marker, the full run re-placed it and reached the boiler in 51 versus 491 steps on task 40 and 93 versus 436 on task 44.The marker had gone unused for nineteen tasks; on task 34, following its stored route added 11 steps relative to no-consolidation.Source: §D.2.1 · p. 39, Figure 18 · p. 39
  • In House C, the calendar marker’s map coordinates changed across days despite simulator positions within 0.5 m; the existing label sufficed on task 40, but misled the run on task 42.The full run re-placed the marker across four tasks, while no-consolidation named the place afresh; re-placement can restore current-frame guidance, but the label can still point toward the wrong goal.Source: §D.2.1 · pp. 39–41, Figure 19 · p. 40
  • For House C’s piano-corner label, placements narrowed from 1.64 to 1.30 to 0.00 m from the goal, and the full run averaged 74 steps across seven piano tasks versus 122.The full run failed task 72 while no-consolidation succeeded, but both succeeded on tasks 74 and 78; marker updates did not guarantee a shorter search on every visit.Source: Figure 21 · p. 41, §D.2.1 · pp. 41–42

D.2.2 Recording failure and checking completion (Cases 5 to 7)

These cases examine how retained and consolidated memories shape later searches, including how notes, markers, procedures, and detailed records can support navigation and recovery. They also show how inherited advice or mistaken completion judgments can mislead, underscoring the need to check memories against new evidence and preserve the conditions behind them.

D.2.2 Recording failure and checking completion (Cases 5 to 7)

  • A dresser note connecting the repeated wording to a basement-level target let the full run resolve its floor and target before route planning, unlike repeated upstairs searches without consolidation.An observation from task 43 was carried through the task account and consolidation into a skill; the no-consolidation run only recorded the answer after 493 steps and a recovery.Source: §D.2.2 · p. 42
  • 18% of House B tasks closed within ten steps with the full mechanism, versus 8% without; at task 62 the verifier accepted the wrong-closet closure after four steps.The full session instead used a within-day marker to close at the correct closet with zero steps, despite a conflicting skill note.Source: §D.2.2 · p. 43, Figure 23 · p. 43
  • The checker accepted false completions on 10% of House A tasks without consolidation and 3% with it, and a certified mistaken dresser identification entered memory and misdirected a later search.A later “unknown” verdict did not automatically retract the earlier house-note account.Source: §D.2.2 · p. 44

D.2.3 Instruments and procedure (Cases 8 and 9)

  • The full run reused a stored response to recurring tracking faults, recording seven recovery requests versus thirteen without consolidation, but a retained remedy sometimes lost its causal explanation.In House B, the frame-change fault was diagnosed three times, yet the saved note kept the remedy rather than the cause, turning a frame-reset warning into general distrust of route previews.Source: §D.2.3 · pp. 45–46

D.2.4 Auditing and consolidating memory (Cases 10 and 11)

  • At task 36, the full run recognized a dark refrigerator image as another view of the known appliance and finished in 36 steps, while the no-consolidation run searched for a second appliance and failed after 490.The consolidated open question prompted additional views, and its answer survived subsequent summaries; the traces do not attribute the step difference to any single memory component.Source: Figure 25 · p. 46, §D.2.4 · p. 47
  • Across nine boiler tasks, the full run succeeded on all nine at 107 steps per task, while the no-consolidation run succeeded on 89% at 274 steps per task.The full run combined a remembered landmark route with current markers; source references supported auditing rule revisions, though uncertain verifier output could also replace a useful rule with an incorrect one.Source: §D.2.4 · p. 47

D.3 When stored experience is declined (Cases 12 to 14)

  • In Case 12, the full run followed a remembered landmark route despite initially distrusting an accurate marker, then corrected its reason on arrival; it reached the boiler in 94 steps versus 317 without consolidation.Both runs succeeded; the trace does not establish that rejecting the marker had zero cost relative to using it with a valid route.Source: §D.3 · p. 48
  • In Case 13, the full run succeeded in 123 steps after a goal photograph contradicted a consolidated plant-avoidance rule, while the no-consolidation run failed; Case 14 concerns repairing a bad entry from its consequences and journal.Source: §D.3 · pp. 47–48
  • The session keeps the room description and photograph useful but rejects an inherited instruction to avoid the plant when the new goal asks for it, recording the conflict for consolidation to revise.This illustrates selective reuse: room information remains useful even when its associated instruction does not fit the current goal.Source: §D.3 · p. 49
  • Simulator checks found that inherited-item doubts were often right in effect but attributed to the wrong cause; House B had 21 checkable judgements holding in effect, and House C had 8 of 17 holding.House A had nine of 17 marker doubts right, four right in effect with the wrong cause, and four wrong; House C also had seven holding with the wrong cause and two contradicted. Only written judgements were counted.Source: §D.3 · p. 49, Figure 26 · p. 50
  • The cases identify five memory-reuse risks, including stale coordinates after a frame reset, false completion recorded as fact, failed-search leads ending in dead ends, and overgeneralized prohibitions.Records also show that note content, within-day markers, and the ledger can guide decisions, while the ledger can carry a wrong observation into a later note.Source: §D.3 · p. 50

D.4 Archive construction and access

  • Consolidation with map and task-record retention fixed raises pooled s-SR 72.8% to 80.5% (+7.7 points; 95% CI [+6.3, +9.1]) and SPL 36.5 to 44.3 (+7.8; 95% CI [+6.5, +9.2]).House notes improve later navigation, and the archive retains successful routes, failed searches, and open questions alongside their sources.Source: §D.4 · p. 51

E Limitations and Future Directions

The evaluations establish limits on NavHarness’s scope: deployments cover finite sequences in static simulated homes, while real-world operation, changing environments, and longer deployments remain untested. Future work includes faster local control, richer spatial memory, selective archive access, and alternative supervision and verification designs.

E Limitations and Future Directions

  • Continuous-deployment evidence covers finite task sequences in static simulated houses, and case studies explain decisions rather than estimate how often mechanisms occur across homes.Paired sequences can diverge in starting pose; route analysis additionally compares tasks completed by both runs from nearby starts.Source: E Limitations and Future Directions · p. 51

E.1 Ownership of the control loop

  • Repeated model decisions and tool interactions add latency; a faster local controller could own rollouts while the larger model manages goals, memory, and task boundaries.In deployments A–C, the full system takes 38–45 model turns per task on average; turns, tool calls, and movement steps are distinct.Source: §E.1 · p. 51, §E.2 · p. 51

E.3 Scope of real-world evaluation

  • Evaluation is simulation-only; physical deployment also needs robust communication, low-latency obstacle avoidance, independently measured trajectories, and substantial real-world outcome checking.Sensor-based SLAM removes dependence on simulator pose, but does not address those requirements.Source: §E.3 · p. 51

E.4 Spatial representation

  • The ORB-SLAM3 spatial state cannot represent object motion, temporal change, or manipulable-object state, and map drift or misalignment can persist.Loop-closure corrections are not propagated into prior occupancy cells; richer map representations would require adapting map-query and update tools, while the harness does not require a particular representation.Source: §E.4 · pp. 51–52

E.5 Archive access over longer deployments

  • As experience accumulates, selective retrieval and revision of stale or conflicting accounts become increasingly important; the current design exposes house overviews and recalls individual notes on demand.Consolidation organizes task experience into house, room, and skill notes with source references.Source: §E.5 · p. 52

E.6 Static environments and deployment horizon

  • Static scanned environments and finite task sequences do not test adaptation to moved objects, changing layouts, or people, and record revision does not establish environmental change tracking.Evaluating memory revision under environmental change and over longer deployments remains future work.Source: §E.6 · p. 52

E.7 Capacity of smaller models

  • With Qwen3.8-27B held fixed, NavHarness raises GOAT task success from 41.4% to 71.7%, but smaller-model adaptation remains future work.Qwen3.5-4B and Qwen3.5-9B struggled with multi-round multimodal navigation; the current study does not update model parameters, though validated process records could support future training.Source: §E.7 · p. 52

E.9 Simulation and model-based judgement

  • Verification checks the first stopping request and completed rollout; the learned judge still accepts false claims on 5.2% and rejects true claims on 4.1% of claimed completions.Frequent checks may interrupt work and multiply judge calls, while continuous supervision requires a full-duplex interface and low-latency verifier unavailable in the present transports.Source: §E.9 · p. 52, §E.8 · p. 52

Related papers

Loading 2609.34276v1…