MOON: Learning When to Forget and How to Move On in Vision-and-Language Navigation

1 Introduction

• Clearing raises pooled success from 13.1% to 26.3% across 779 delayed states, while retaining only history up to deviation onset raises it to 16.3%.Removing later off-route frames recovers much less of the harm than clearing the history.Source: Figure 1 · p. 2, §1 · p. 2
• The paper asks how visual history affects recovery because prior recovery methods changed policies and evaluated mixed on-route and off-route episodes, leaving history’s role during recovery unisolated.An incorrect action can take an agent off the instructed route, where subsequent errors may compound into task failure.Source: §1 · p. 2
• Clearing history raises success from 13.7% to 28.6% at deviation onset, but lowers it from 76.0% to 48.3% on route across four policies.Both effects hold for every policy.Source: §1 · p. 2
• Retained history can keep a policy from turning when the new state calls for one, a failure mode the paper calls behavioral inertia.After clearing, at least one of eight sampled rollouts succeeds in 61.6% of states where the retained greedy rollout fails, versus 30.3% with history retained.Source: §1 · pp. 2–3
• MOON adds a deviation judge that clears history when it predicts the agent is off route and a GRPO-trained recovery policy to a frozen policy, adding less than 1% of its parameters.The judge avoids clearing on route, where clearing is harmful.Source: §1 · p. 3
• MOON reaches the highest R2R-CE success rate among compared methods, 2.1 points above Qwen-RobotNav-4B, and raises its frozen base by 5.2 R2R-CE and 6.0 RxR-CE points.It also improves zero-shot open-vocabulary object-goal navigation on HM3D-OVON without retraining.Source: §1 · p. 3

2.1 Preliminaries

• The analysis uses paired history interventions from the same simulator state, with deviation defined as distance exceeding 2 m from the reference route.It uses 600 randomly sampled R2R-CE val-unseen episodes and four frozen policies.Source: §2.1 · p. 3, §2.2 · p. 4

2.2 Two Opposing Roles of Visual History

• Across 915 paired onset states, clearing raised pooled success from 13.7% to 28.6%, while on 1,465 on-route states it lowered success from 76.0% to 48.3%.The pattern held for every policy in both conditions.Source: §2.2 · p. 4
• For on-route states, pooled success rose with history length, reaching 27.7 points above clearing with full history; after deviation, the deficit widened to 13.2 points.At k = 2, the corresponding differences were 9.3 points above clearing on route and a 3.5-point deficit after deviation.Source: §2.2 · p. 4
• At 779 delayed states, retaining history only through deviation onset raised pooled success from 13.1% to 16.3%, but left a 10.0-point deficit.The deficit persisted when history ended before its first frame beyond 0.5 m from the route.Source: §2.2 · p. 4

2.3 Behavioral Inertia

• After deviation, shuffled or cross-scene histories retained nearly all of the 13.2-point deficit, while both manipulations lowered success on route for every policy.Shuffling and swapping left 13.0 and 12.5 points of the deficit, respectively.Source: §2.3 · p. 4
• After a deviation, policies with retained history continued without turning in 48.1% of onset states and 46.8% of delayed states, versus 1.9% and 4.8% for the expert.Clearing reduced policy continuation to 10.9% and 9.3%; the authors call the retained-history effect behavioral inertia.Source: §2.3 · p. 5

2.4 Beyond Clearing

• Even an oracle choosing whether to clear left 65.0% of onset states unresolved; clearing rescued 195 retained-rollout failures but broke 58 retained-rollout successes.Source: §2.4 · p. 5
• With eight samples, pooled coverage on onset states with failed retained greedy rollouts rose from 30.3% with retained history to 61.6% after clearing, but 38.4% still had no successful cleared rollout.This offline coverage assumes a success can be recognized among candidates.Source: §2.4 · p. 5

3 Method

• MOON equips the frozen policy with a judge and recovery policy, both using LoRA adapters; an alarm clears visual history before recovery takes over.At deployment, the judge raises an alarm when the deviation score exceeds its conformal threshold, and recovery proceeds in one greedy rollout.Source: §3 · p. 5, Figure 4 · p. 6

3.1 Deviation Judge

• The deviation judge uses the instruction, current observation, history, and proposed action chunk to predict sustained deviation, training on LOST and OK labels.A LOST label requires being off route now and at half or more of the later steps within a fixed horizon; stop actions and chunks containing a stop are not judged.Source: §3.1 · pp. 5–6
• A moving average of recent log-odds margins suppresses single-point noise; a split-conformal threshold sets the alarm that clears history and hands control to recovery.Clearing preserves the instruction and current frame, happens at most once per episode, and the recovery policy then acts greedily while history re-accumulates.Source: §3.1 · pp. 6–7
• The threshold is calibrated on completed held-out episodes so that, under exchangeability, the chance of interrupting an episode the base policy would complete is at most α.The alarm uses information available to the agent and can precede the deviation time defined in the paper.Source: §3.1 · p. 6

3.2 Recovery Policy

• From first-alarm states, GRPO samples recovery rollouts, scores only their post-alarm suffixes with suffix-SPL, and updates only LoRA ψ using group-relative advantages.The reward ranks successes above failures and rewards shorter successful paths; groups with equal rewards are discarded, and clearing leaves fewer such groups than retaining history.Source: §3.2 · p. 7

4.1 Experimental Setup

• MOON is evaluated on full val-unseen splits of 1,839 R2R-CE and 3,669 RxR-CE episodes, with a frozen NaVIDA-Qwen3VL base and adapters trained on the training splits.The judge threshold is calibrated per dataset; decoding is greedy and paired outcome comparisons use the exact McNemar test.Source: §4.1 · p. 8

4.2 Main Results

• MOON achieves 69.0 SR on R2R-CE, 2.1 points above Qwen-RobotNav-4B, and raises its frozen base’s SR by 5.2 points on R2R-CE and 6.0 on RxR-CE.On RxR-CE, its SR and SPL are within 2.7 points of the best method, Qwen-RobotNav-8B.Source: §4.2 · p. 8
• On R2R-CE, the system rescues 24.6% of base failures and breaks 5.8% of base successes; on RxR-CE, it rescues 24.7% and breaks 4.1%.Figure 5 shows alarm shares among base failures or successes, colored rescue or break shares, and their difference rather than an SR change.Source: §4.2 · pp. 8–9, Figure 5 · p. 9
• Without retraining, MOON raises HM3D-OVON zero-shot SR from 40.7 to 57.2 and SPL from 21.7 to 35.4, the best among compared methods.Source: §4.2 · p. 9

4.3 Ablation Study

• Without the judge, SR falls to 59.9; without the recovery policy, clearing still reaches 65.6, below the full system.Taking over retained history without clearing reaches 67.4, not significantly below the full system; Table 2 covers core components.Source: §4.3 · p. 9, Table 2 · p. 9
• A random trigger lowers SR below the base, while GRPO raises greedy recovery success from 35.1% to 47.1% on alarmed episodes.On the first training iteration, pass@8 is 50.8% with cleared rollouts versus 22.7% with retained history; the two runs end at similar closed-loop SR.Source: §4.3 · pp. 9–10, Figure 6 · p. 9, Table 3 · p. 9

• With 0.81% additional parameters, MOON raises frozen-base success by 5.2 points on R2R-CE and 6.0 on RxR-CE, while improving zero-shot object-goal navigation without retraining.Source: §5 · p. 10, Table 5 · p. 10

6 Limitations

• MOON is deployed on NaVIDA-Qwen3VL, while the analysis covers four policies; evaluation on newer open-source history-conditioned policies remains a natural next step.MOON trains two lightweight LoRA adapters while keeping the policy frozen.Source: §6 · p. 10
• Alarms cover 56.3% of episodes the base fails on R2R-CE and arrive a median 13 atomic steps after deviation onset on R2R-CE and 21 on RxR-CE.The authors state that earlier, more accurate detection could recover more episodes; a ground-truth trigger reaches SR 71.2 against 69.0 (Table 3).Source: §6 · p. 10
• Clearing breaks 5.8% of episodes the base completes on R2R-CE and 4.1% on RxR-CE, with each episode cleared at most once.Source: §6 · p. 10

A.1 Vision-and-Language Navigation and Visual Memory

• Prior visual-memory work often retrains policies or evaluates full episodes, whereas this study holds policies fixed and separates on-route from post-deviation states.The closest prior detector leaves the VLA context untouched, while MOON clears history and hands control to a recovery adapter at the current pose.Source: §A.1 · p. 15, §A.2 · p. 15

A.3 Corrective Training and Reinforcement Learning

• Corrective training and reinforcement-learning evaluations do not isolate visual-history effects on recovery under a fixed policy; MOON instead clears history at deployment and trains recovery from alarm states.Prior methods include synthesized corrective supervision, geodesic-deviation supervision, and task-reward reinforcement learning.Source: §A.3 · p. 15

A.4 History Dependence in Imitation Learning

• After a deviation, retained visual history can keep a navigation policy acting on expert-history correlates that no longer hold, even as the current state calls for a course change.MOON clears history only after an alarm, rather than reshaping training or replacing history throughout an episode.Source: §A.4 · pp. 15–16

A.5 Calibrated Thresholds and Reinforcement Fine-Tuning

• MOON calibrates a judge threshold on episodes completed by the frozen policy and trains a LoRA recovery adapter with GRPO from cleared alarm states.GRPO normalizes rewards within sampled-output groups without a learned critic; the reward applies to the rollout after the alarm.Source: §A.5 · p. 16

B.1 State Populations and Uncertainty

• The analysis uses fixed-policy interventions on 600 R2R-CE val-unseen episodes; pooled uncertainty resamples policy states independently, and the 95% intervals are pointwise, not simultaneous.The delayed population contains 779 states, with 749 (96.1%) still more than 2 m from the route; populations hold at most one state per episode within each policy.Source: §B.1 · p. 16, Table 6 · p. 16

B.2 History Constructions and Comparators

• Prefix + current retains the history through deviation onset and the current observation; route-only history is defined for 420 eligible states, and the expert-plus-tail construction can retain off-route content.The 1.5, 1.0, and 0.5 m prefixes leave pooled deficits of 9.9, 8.2, and 7.4 points.Source: §B.2 · p. 17, Table 7 · p. 18

B.3 History Substitution

• At 779 delayed states, shuffled and swapped histories leave deficits of 13.0 and 12.5 points, close to 13.2 points for full-length retained history; on route, substitutions lower success by 5–20 points.Copies of the current frame and token-matched noise still leave deficits of 8.7 and 7.7 points.Source: §B.3 · p. 18

B.4 Action Continuation

• Continuation is scored after forward steps; treating stop-ending chunks without a turn as non-continuing lowers retained-history rates from 48.1% to 44.9% at onset and 46.8% to 41.8% at delayed states.Shuffled and swapped histories differ from retained-history continuation estimates by less than 4 points pooled, and remain above clearing and the expert for every policy and population.Source: §B.4 · pp. 18–19, Table 8 · p. 19

B.5 Greedy Selection and Sampling Coverage

• The greedy selection oracle reaches 35.0% success on 915 onset states, 6.3 points above always clearing, but bounds only selection between two frozen greedy continuations.Offline pass@k coverage assumes successful samples can be recognized and is not a deployment success rate.Source: §B.5 · p. 19, Table 9 · p. 20

C.1 Base Policy

• The frozen NaVIDA-Qwen3VL base follows the official NaVIDA training recipe with Qwen3-VL-4B-Instruct as its backbone.Table 10 lists its configuration.Source: §C.1 · p. 20, Table 10 · p. 20

C.2 Judge

• The judge scores eligible decision points using up to nine history frames, the instruction, recent actions, and the proposed action chunk, then answers with one token.Its training uses rollouts from R2R-CE, RxR-CE, and ScaleVLN; the adapter is LoRA rank 32, and checkpoint step 1,500 is selected.Source: §C.2 · p. 20

C.3 Calibration

• Calibration assigns each dataset its own threshold from the maximum smoothed judge score over eligible decisions in completed greedy episodes.The selected levels are α = 10% for R2R-CE and α = 7.5% for RxR-CE; forced stops can exceed the stated bound by at most their probability among completed episodes.Source: §C.3 · pp. 20–21

C.4 Training Start States

• The pool contains 5,510 first-alarm states from R2R-CE, RxR-CE, and ScaleVLN rollouts, with 24.2% from completed episodes and 25.8% within the route-distance threshold.Training rollouts begin at the recorded pose with the remaining step budget and current frame alone; held-out-path episodes are excluded.Source: §C.4 · p. 21

C.5 Counterfactual Stop Advantage

• Removing the counterfactual stop term has no measurable effect on success or stopping behavior in the reported ablation.At iterations 3 and 9, success rates are 66.3 versus 66.2 and 66.6 versus 67.0 with the term, respectively.Source: §C.5 · p. 21

C.6 Recovery Objective

• The recovery objective combines clipped PPO losses for continued chunks and stop choices with a KL estimate that regularizes the adapted policy against the frozen policy.The stop reward has a closed form because stopping ends the episode; the stop and continuation advantages are split across the two choices.Source: §C.6 · p. 21, §C.5 · p. 21

C.7 Training and Decoding

• Each iteration draws 128 alarm states and samples eight rollouts per state at temperature 1.6 with top-p = 1.The reward uses λ = 0.3 and the environment’s 3 m success radius.Source: §C.7 · p. 22
• The zero-initialized rank-16 LoRA adapter starts identical to the base policy, while decoding remains greedy before and after alarms.The adapter is applied to language-model projections, four vision blocks, and the vision merger; evaluation and deployment both use greedy decoding.Source: §C.7 · p. 22

D.1 Training-Side Ablation over Iterations

• On 527 alarmed R2R-CE episodes, greedy success reaches 47.1% at iteration 15, while GRPOstart without a judge stays at the base level.A binary success reward reaches 66.5 at iteration 9 versus 67.0 with suffix-SPL; Table 11 gives success rates across iterations and training-start conditions.Source: §D.1 · p. 22, Table 11 · p. 22

D.2 Judge Counts

• On R2R-CE, the recovery policy raises the share of alarmed failures it completes from 29.1% with clearing alone to 43.7%, while lowering the share of alarmed successes it breaks from 50.0% to 44.7%.The other 43.7% of R2R-CE failures never receive an alarm, bounding potential gains; calibration on held-out paths within training scenes can yield false-alarm rates above nominal levels.Source: §D.2 · p. 22, Table 12 · p. 23

D.3 Alarm Level

• Every tested alarm-calibration level improves on the base, and the reported levels give the highest success on both datasets.Ten of eleven leave-one-scene-out folds select the reported level; false-alarm rates exceed α by at most 3.0 points on R2R-CE and 0.6 on RxR-CE.Source: §D.3 · p. 23, Table 13 · p. 23

D.4 Judge Query Interval

• Querying the judge at every second or fourth eligible decision point raises about as many alarms but later, with success rates falling to 67.9 and 67.0.Table 14 compares judge query intervals on R2R-CE; the deployed interval is marked in the table caption.Source: §D.4 · p. 23, Table 14 · p. 23

D.5 Zero-Shot Object Navigation

• The zero-shot HM3D-OVON evaluation uses unchanged adapters and an R2R-CE-calibrated threshold, with the base scoring 40.7 SR and 21.7 SPL under the same protocol.Because HM3D-OVON has no route labels, an alarm marks a score above the R2R-CE threshold rather than a measured deviation; the R2R-CE calibration bound does not carry over.Source: §D.5 · pp. 23–24, Table 15 · p. 24

E Qualitative Examples

• The displayed supplementary episodes were hand-selected from cases where the frozen policy fails and MOON completes, so they illustrate recoveries rather than typical outcomes.The collection includes three supplementary episodes per dataset, while Table 12 reports how often alarms rescue or break episodes.Source: E Qualitative Examples · p. 25, Figure 8 · p. 25, Figure 11 · p. 28

F.1 Navigation Policy

• The navigation prompt supplies up to eight uniformly sampled earlier frames and the current frame, with run-time inputs and frame tokens formatted as specified.After a history clearing, only the current frame is retained; on HM3D-OVON the instruction names the goal category.Source: §F.1 · p. 31, F Prompts · p. 31

F.2 Deviation Judge

• The deviation-judge prompt includes visual history, the instruction, recent executed actions, and the proposed action chunk, and requests an OK, lost, or amb response.The judge is supervised on the answer token, with lost for LOST and OK for OK.Source: §F.2 · p. 31