Pith.

Robotics · arXiv:2610.39446 v1 · submitted 29 Oct 2026 · 86 references

Ask to Evolve: A Unified Harness for Lifelong Embodied Navigation with Hybrid Action Space and Persistent Memory

Kai Sheng, Liuyi Wang, Jinlong Li, Haojie Dai, Chengju Liu, Qijun Chen

The paper says

“Giving embodied agents the freedom to act autonomously, initiate interactions, and understand their surroundings through experience is a step toward bringing them to life in the physical world.”

more of the abstract

Realizing this vision requires action, interaction, and memory to work together, so that each encounter can inform current decisions and contribute experience for future tasks. We present EvoLife, a unified navigation harness that enables a multimodal embodied agent to choose how to move, when to ask, and what experience to carry forward. Its hybrid action space supports waypoint navigation, local discrete adjustments, pixel-based navigation, and object-directed navigation. Active interaction allows the agent to seek clarification or assistance when it encounters ambiguity or difficulty, while persistent spatial memory binds observations and dialogue to places, retaining useful experience to guide subsequent visits. We evaluate eight benchmarks across four task families: vision-language navigation, object navigation, interactive instance navigation, and continuous multi-task navigation. With an average of only 1.87 active interactions per episode, our method achieves a first-visit success rate (SR) of 100% across four 100-episode sets spanning vision-language navigation and object navigation. Complete-sequence SR reaches 98% on 50 LHPR-VLN sequences and 94% on 50 GOAT-Bench sequences. In the R2R revisit setting, success weighted by path length (SPL) increases from 82.23% to 93.31% with retained experience. Multi-day physical navigation further demonstrates exploration, proactive help-seeking, and revisiting remembered places. Together, these results advance the vision of bringing embodied agents to life through autonomous action, proactive interaction, and the accumulation and reuse of experience.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 30 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. The harness is real and the paper is honest about its protocol, but the 100% SR headline is an oracle ceiling: the Sec 3.4 guide holds pose, target, and shortest path, and the abstract never conditions the claim on it. This line is wrong

4 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 86 references say.

The paper says, at Abstract, §4.2, §3.4 (Eq. 4), §5

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The headline claims (100% SR on four 100-episode sets; 98%/94% complete-sequence SR [b6, b51, b239-b243]) are never conditioned, in the abstract or in §4.2, on the online guide of §3.4, which receives the agent's current pose, target location, shortest navigable path, relevant map data, and, for CoIN/VL-LN, the target instance identity (Eq. 4).

more of objection 1

The channel is disclosed in §3.4 and Fig. 3, but §5's limitation list contains only latency and API cost [b347-b349], and the paper states no scope condition on the headline. This is load-bearing: the paper's own ablations attribute most of the margin to the answer-key channel. Table 5 shows R2R-CE SR rising from 86% to 100% when interaction is enabled; Table 7 shows HM3D-OVON SR falling from 100% to 69% when arrival-confirmation requests are removed; and Table 4 shows the same-backbone NavHarness at 83.7% s-SR versus 99.3% for EvoLife, bracketing roughly 16 points of the GOAT gap. The abstract must state that the headline results require a guide with privileged spatial information, and the Fig. 3 claim that the setup 'simulates natural human-to-human interaction' should be softened, since a human helper without pose and shortest-path data cannot reproduce this channel.

The paper says, at §4.3 (Table 5), §4.5 (Table 7)

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

No-interaction controls exist only for R2R-CE; for RxR-CE, HM3D v2, HM3D-OVON, VL-LN, CoIN, LHPR-VLN, and GOAT-Bench, the headline numbers are never decomposed into the contribution of the privileged guide, so the 'outperforming SOTA by 7.0-57.0 points' statements in §4.2 [b240-b249] are not attributable to any specific component.

more of objection 2

Table 7 covers only four of the eight benchmarks, and it removes only arrival-confirmation requests rather than the direction and target-identity guidance that the guide also supplies (as in the Fig. S4 and Fig. S5 examples). The revision should add no-interaction controls (or no-arrival-confirmation controls) for all benchmarks, or explicitly state which claims cannot be decomposed; without this, the reader cannot separate harness competence from oracle assistance on seven of the eight benchmarks.

The paper says, at §4.1, §4.2 (Tables 3-4)

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

Only R2R-CE uses the fixed 100-episode OpenNav evaluation set; the other benchmarks use 100 or 50 randomly selected episodes, and the published SOTA numbers quoted in Tables 3-4 were computed on different (typically larger) evaluation subsets. The absolute SR gaps reported in §4.2 (e.g., 13.0 points over ASENA on R2R-CE, 31.7 points over SuperNav on HM3D-OVON) are therefore not controlled comparisons.

more of objection 3

Additionally, 100% success on 100 episodes has a 95% confidence interval of roughly [96.4%, 100%], so '100%' is not a precise estimate of the underlying success rate; I ask for the selection seeds, a re-evaluation of the strongest baselines on the same episode subsets where feasible, and interval or variance reporting before the superiority claims are stated as absolute differences.

The paper says, at §4.3 (Table 5)

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The revisit claim (SPL improving from 82.2% to 93.3% with retained experience [b8, b53]) inherits the oracle channel: the place records and dialogue stored during the interactive first visit were produced with answers from a guide that holds the target location and shortest path, so the remembered content is oracle-derived.

more of objection 4

The '× ✓' memory-revisit row in Table 5 has no corresponding 'non-interactive first visit followed by memory revisit' condition, which means the memory benefit is not isolated from the privileged information that populated the memory. I ask for a revisit condition that follows a non-interactive first visit, or an explicit statement, in the limitations, that the reported memory gain applies only to experience acquired through privileged guidance.

6 minor objections

The paper says, at Table 4

Pith says: minor objection 1 of 6

The column header 'GOA T-Bench' contains a typo; it should read 'GOAT-Bench'.

The paper says, at Fig. 2

Pith says: minor objection 2 of 6

The task text in the figure reads 'stop by the dinning room table'; 'dinning' should be 'dining'.

show 4 more

The paper says, at §4.5

Pith says: minor objection 3 of 6

The procedure 'directly executes STOP upon the first arrival-confirmation-related request' requires a definition of what counts as an arrival-confirmation-related request; if this classification was done manually or by a rule on the tool-call logs, that should be stated.

The paper says, at §4.7

Pith says: minor objection 4 of 6

The failure analysis says the agent 'successfully identifies the correct target in every case,' but the total number of recorded failures is small (at most one VL-LN episode, one LHPR sequence, and three GOAT sequences under the reported SR figures); the claim should be presented with the failure count and treated as anecdotal rather than as evidence that target recognition is solved.

The paper says, at §4.8

Pith says: minor objection 5 of 6

The real-world section asserts task success without an independent success criterion; please state how arrival was verified (e.g., manual judgment vs. measured pose error) and report approximate localization error over the 166 m Day-1 route.

The paper says, at §3.2 (Eq. 1)

Pith says: minor objection 6 of 6

The policy parameters θ in Eq. (1) are never defined; since EvoLife is a training-free harness, the text should state explicitly that θ denotes a fixed backbone prompt/tool-selection configuration and that no parameters are learned.

What would settle it

Re-run the identical harness on the same 100 R2R-CE episodes (plus RxR-CE and HM3D-OVON if feasible) with ask_human active but the guide restricted to the navigation instruction and current observations only — no pose, no target, no shortest path, no map, no instance identity — using the same GPT-6 Astra backbone. Compare SR/SPL against the three existing anchors: no-interaction 86.0/63.7 (Table 5), no-arrival-confirmation 90.0/74.0 (Table 7), and full oracle 100/82.2. If the observation-only guide lands near 86–90 SR, the answer-key channel accounts for the remaining margin and the abstract's unconditional 100% must be re-scoped. If it sustains approximately 95+ SR, the interaction design carries its weight with human-grade help and the concern softens.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 30 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 1 min 37 s · 31,117 tokens read, 14,179 written
  3. Editor wrote the verdict above 27 s · 32,248 tokens read, 4,664 written
  4. Referee major revision: 4 major objections, 6 minor 1 min 1 s · 31,479 tokens read, 8,467 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The headline result is genuine only under the paper's own protocol. The guide in Sec 3.4 (Eq. 4) is a separate model that answers the agent's questions with privileged information: the agent's current pose, the target location, the shortest navigable path, relevant map data, and for instance tasks the target's identity. That is an answer-key channel. The paper discloses it in Sec 3.4, but the abstract's unqualified 100% SR claim (b6) and the margins over SOTA in Tables 3-4 never state it as a scope condition. As presented, those margins measure the harness plus a guide that can read the solution, not navigation skill.

That said, the paper is better than that summary makes it sound, and I want to give the authors credit. The integrated system — four movement modes, proactive asking, place-grounded dialogue memory in one decision loop — is a legitimate new configuration that I don't think appears in NavHarness, HarnessVLN, or the other cited harnesses. The ablations are unusually informative: the no-interaction R2R-CE control lands at 86% SR, right at published SOTA, which gives a clean decomposition of what interaction adds. The beyond-arrival-confirmation analysis (Table 7) directly quantifies the oracle's contribution, and the failure analysis shows every failure was a stopping-distance miss, not target recognition. The R2R revisit numbers (82.23 to 93.31 SPL) are plausible as a memory-reuse effect. The paper is honest about what it did; the soft spot is how it presents it.

Where it's soft, in proportion: the abstract and comparison sections never condition the headline on the oracle channel, and Sec. 5's limitations mention only latency and cost (b347-b349). Table 7 shows how much this matters — HM3D-OVON drops from 100% to 69%, R2R from 100% to 90%, without arrival-confirmation questions. The GOAT-Bench table gives the same lesson from the baseline side: NavHarness with the same backbone but no oracle scores 83.7% s-SR against 99.3%. No-interaction controls exist only for R2R-CE, so for seven of eight benchmarks the numbers are never decomposed. The memory-reuse claim inherits the issue, since first-visit dialogue is stored in place records and dialogue is never ablated from memory. And there is no code, data, seeds, or prompts; the backbones are closed proprietary models, so the numbers are not independently checkable.

Who this is for: people working on LLM harnesses and interactive navigation evaluation. The system design deserves a serious referee, and the paper is worth engaging once the framing is fixed — state the oracle condition in the abstract, report every benchmark with and without interaction, and release the harness.

The strongest objection

The central claim — 100% SR on four 100-episode sets at 1.87 interactions per episode (abstract, Sec 4.2) — is conditioned nowhere in the abstract on the online guide of Sec 3.4 (Eq. 4), which receives the agent's current pose, the target location, the shortest navigable path, relevant map data, and, for CoIN/VL-LN, the identity of the target instance. The paper discloses the channel in Sec 3.4, but its Sec 5 limitations cover only latency and API cost; the answer-key dependency is never stated as a scope condition on the headline. The paper's own ablations show the channel is the source of most of the margin. On R2R-CE the no-interaction control (Table 5) is 86% SR, essentially ASENA's 87%; interaction adds 14 points, and Table 7 attributes 10 of them to arrival-confirmation questions answered by a guide that knows the agent's pose and target — an oracle-verified stop decision, which is precisely the subproblem VLN success metrics exist to test. On HM3D-OVON the same ablation costs 31 points (100% to 69%). No-interaction controls exist only for R2R-CE; for RxR, HM3D v2, HM3D-OVON, VL-LN, CoIN, LHPR-VLN, and GOAT-Bench the headline numbers are never decomposed. The one cross-benchmark decomposition available, GOAT-Bench in Table 4, is telling: same-backbone NavHarness without the oracle channel scores 83.7% s-SR versus 99.3%, bracketing the channel's contribution at roughly 16 points. The revisit claim (SPL 82.23% to 93.31%) inherits the same issue because place records store first-visit oracle dialogue, and dialogue is never ablated from memory. Under its stated protocol the system likely does what it claims; the problem is that the central claim, as presented in the abstract and comparison sections, concerns a protocol that no baseline shares.

I read the paper as a systems contribution: a harness (hybrid actions, proactive questions, persistent memory) wrapped around a proprietary VLM, evaluated in Habitat with a GPT-6 Astra backbone. The central claim is conditional in fact but not in presentation: Sec 3.4 discloses that the guide holds the agent's pose, target, shortest navigable path, map, and (on instance tasks) the target instance identity, yet the abstract states the 100% figures without that condition, and the comparison sections present 13–31.7 point margins over baselines with no such channel. The strongest internal evidence that this is the decisive factor is the paper's own table structure. Table 5's no-interaction R2R row (86% SR, 63.7 SPL) sits within noise of ASENA's 87%, so the harness alone is roughly SOTA, not superhuman. Table 7 shows that removing arrival-confirmation questions — answered by a guide that knows the pose and target, i.e., an oracle-verified stop decision — costs 10 points on R2R (100 to 90) and 31 points on HM3D-OVON (100 to 69). Because VLN success metrics exist precisely to test when an agent knows it has arrived, having the oracle answer 'are we there yet?' converts the hardest subproblem into a dialogue-following exercise. Table 4's GOAT-Bench row gives the only cross-benchmark decomposition: NavHarness, same GPT-6 Astra backbone, no oracle channel, 83.7% s-SR versus 99.3% with it. What remains genuinely unresolved is the size of the non-arrival interaction benefit under human-grade information: Table 7 preserves directional and target-clarification questions, but those answers still come from a guide holding the target and shortest path. The paper never runs the channel with an information-restricted guide, and never runs no-interaction controls on the other seven benchmarks. It also never ablates dialogue out of the place records used in the R2R revisit, so the 93.31 SPL claim is entangled with stored oracle answers. None of this indicts the authors' honesty — the ablations are more transparent than typical for this literature, and the no-interaction R2R floor at 86% gives the harness real credit. It does mean the abstract-level claim needs to be re-scoped to 'with an answer-key guide,' and the SOTA margin comparisons need the same condition attached. That is exactly what CONDITIONAL should carry: accept the framework and methodology, require re-scoped claims. The observation-only-guide control would settle sharply whether the interaction design contributes beyond executing oracle directions.

The referee's report

If the results are taken as conditional on the privileged guide, the paper is a well-executed systems contribution: the harness is backbone-agnostic (Table 6), it involves no learned parameters or parameter fitting, the no-interaction R2R control (86% SR, 63.7% SPL) sits near published SOTA and shows the harness itself is not harmful, and the ablations in Tables 5 and 7 are unusually informative about where the gains come from. The beyond-arrival-confirmation analysis (Table 7) and the failure analysis (Sec. 4.7) are honest attempts to probe the mechanism rather than mere saturation claims. The weakness is that the paper's own measurements show the answer-key channel, not the harness, is the source of most of the headline margin, and this dependency is not carried into the abstract, the results section, or the limitations discussion. The central claim is therefore defensible only after reframing: the paper reports what a privileged-information interaction channel can add atop a near-SOTA harness, not an unconditional navigation result. Properly scoped, this is a useful reference point for the interactive-navigation community.

Pith's reading of the paper (a machine's summary, not the authors')

EvoLife is a harness for embodied navigation that unifies three capabilities in one decision loop: four ways of moving (waypoint, local step, pixel-selected goal, object-directed approach), proactive questions to a guide, and persistent place memory that binds observations and dialogue to locations. The paper claims this combination is what lifelong navigation needs, and reports that on first visits to unseen environments it achieves 100% success on four 100-episode benchmarks spanning vision-language navigation (R2R-CE, RxR-CE) and object navigation (HM3D v2, HM3D-OVON), with an average of only 1.87 questions per episode. The results are explicitly conditional on an online guide that answers using privileged spatial information — the agent's pose, the target, the shortest path, and map data — so the headline numbers describe the harness plus a well-informed helper, not autonomous navigation alone. The paper also shows that retained memory removes the need to re-ask: on R2R revisits, success weighted by path length rises from 82.23% to 93.31% with interaction disabled, on fewer steps and decisions. Why this matters: if correct, near-saturation success on standard navigation benchmarks no longer requires task-specific training but hinges on giving agents a way to ask, a way to act on the answer, and a memory that makes the answer reusable.

What carries the argument. The central mechanism is the tool-based decision loop in which a multimodal policy chooses among perception, movement, interaction, and memory tools at each step. Movement is expressed through a hybrid action space of four modes — navigate_candidate through geometric waypoints, low-level step for local adjustment, navigate_pixel for image-selected floor goals, and navigate_object for RGB-D localized targets — so the agent can express its intention at the right granularity and inspect the execution feedback. Interaction is the ask_human tool: the agent generates a question with an intent (clarification, guidance, assistance), and a separate multimodal guide returns a natural-language answer (Eq. 4) computed from the instruction, current views, and privileged spatial information including pose, target, shortest path, and map. Memory is organized as named place records (Eq. 5) that tie a pose to descriptions, object labels, historical images, and dialogue, with tools to create, search, inspect, and navigate to places; records outlive individual tasks and are revised by new evidence. The unifying claim is that these three families reinforce each other: an interaction can change the next movement, revise the interpretation of a place, and leave knowledge that a later task reuses without re-asking.

Core claim. On its own terms, the paper's discovery is that "ask to evolve" is a working recipe: an agent that selects among complementary movement tools, decides itself when to ask for clarification or assistance, and stores each exchange as place-grounded memory can navigate unseen indoor environments at ceiling success. Concretely, the harness reports 100% first-visit success on R2R-CE, RxR-CE, HM3D v2, and HM3D-OVON; 99% on VL-LN and 100% on CoIN, where the agent must disambiguate a hidden target instance through dialogue; and 98% and 94% complete-sequence success on LHPR-VLN and GOAT-Bench, where memory persists across ordered subtasks. The interaction that buys these gains is cheap in count — between 1.6 and 3.2 questions per episode with a maximum of ten allowed — and each question is answered by an independent multimodal guide that translates privileged spatial evidence into natural-language directions, landmarks, and spatial relationships. The paper's own ablations delimit what interaction and memory each contribute: without interaction, R2R-CE success falls from 100% to 86%, and removing only arrival-confirmation questions drops HM3D-OVON success from 100% to 69%; with memory but no interaction on a revisit, R2R success stays at 100% while SPL improves to 93.31%. In real-world two-day trials, the same loop handles a closed door by asking a person for help and a changed lighting condition by matching current views to stored place records.

The premise it rests on. The load-bearing premise is that a separate model — the guide — can always answer the agent's questions using privileged information the agent does not have, namely its own position, the target's location, the shortest path to it, and the map layout; the near-perfect results assume this answer channel exists and answers accurately.

If the paper is right (Pith's inferences, not the authors')

If the harness is right, ceiling-level success on standard VLN and object-navigation benchmarks is reachable without task-specific training: a strong general backbone plus the right tool interface suffices, with success differences across backbones (100%, 91%, 98% on R2R-CE) reflecting planner quality rather than harness failure.

A small number of questions buys most of the gain: 1.66 interactions per R2R episode lifts success from 86% to 100%, implying that much of what blocks current agents is resolvable ambiguity rather than missing motor or perceptual skill.

Memory is the cheaper route to the same destination: on R2R revisits the agent keeps 100% success with interaction disabled while raising SPL from 82.23% to 93.31% and cutting motion-equivalent steps from 70.12 to 54.43 and tool decisions from 15.57 to 12.31.

The same interface transfers across four task families — route instructions, object categories, hidden-instance dialogue goals, and ordered multi-stage sequences — and across at least three decision backbones, supporting the claim that action, interaction, and memory can live in one loop.

Interaction works beyond arrival confirmation: with confirmation questions suppressed, success remains at 90–95% on the two VLN benchmarks, and the paper's failure analysis attributes the remaining misses to final stopping precision (for example 0.73 m versus a 0.25 m threshold), not to target recognition.

The guide is effectively an answer key: because it sees the pose, target, shortest path, and map, the 100% figures are best read as an upper bound on what the harness can do with perfect help; the open question is where a realistic helper — a person without map access, or a weaker model — places results between the no-interaction baselines and the ceiling.

Revisit gains inherit the dialogue stored during first visits, so memory quality is downstream of guide quality; a natural extension is to test whether revisits stay near 93% SPL when the stored dialogue comes from a cheaper or less capable guide.

A testable refinement would classify questions by intent and measure each intent's contribution: the arrival-confirmation ablation already isolates one class, and splitting directional guidance from target clarification would show whether "asking to evolve" or "asking to verify" drives the gains.

Because the decision policy is the general backbone rather than a dedicated trained module, a two-model split — a small model deciding when to ask and a strong model planning movement — is a direct route to cutting the per-episode latency and API cost that the paper names as its main limitation.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Giving embodied agents the freedom to act autonomously, initiate interactions, and understand their surroundings through experience is a step toward bringing them to life in the physical world. Realizing this vision requires action, interaction, and memory to work together, so that each encounter can inform current decisions and contribute experience for future tasks. We present EvoLife, a unified navigation harness that enables a multimodal embodied agent to choose how to move, when to ask, and what experience to carry forward. Its hybrid action space supports waypoint navigation, local discrete adjustments, pixel-based navigation, and object-directed navigation. Active interaction allows the agent to seek clarification or assistance when it encounters ambiguity or difficulty, while persistent spatial memory binds observations and dialogue to places, retaining useful experience to guide subsequent visits. We evaluate eight benchmarks across four task families: vision-language navigation, object navigation, interactive instance navigation, and continuous multi-task navigation. With an average of only 1.87 active interactions per episode, our method achieves a first-visit success rate (SR) of 100% across four 100-episode sets spanning vision-language navigation and object navigation. Complete-sequence SR reaches 98% on 50 LHPR-VLN sequences and 94% on 50 GOAT-Bench sequences. In the R2R revisit setting, success weighted by path length (SPL) increases from 82.23% to 93.31% with retained experience. Multi-day physical navigation further demonstrates exploration, proactive help-seeking, and revisiting remembered places. Together, these results advance the vision of bringing embodied agents to life through autonomous action, proactive interaction, and the accumulation and reuse of experience.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-30

The authors' main claim, as the classifier picked it out 1.00

“Together, these results advance the vision of bringing embodied agents to life through autonomous action, proactive interaction, and the accumulation and reuse of experience.”

How the main claim is stated: asserted 0.97

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.07 · human participants 0.09 · controlled comparison 0.50 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.71 · mentions code or data availability 0.04

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Giving embodied agents the freedom to act autonomously, initiate interactions, and understand their surroundings through experience is a step toward bringing them to life in the physical world. background 1.00 0.04 0.06 asserted 0.86
Realizing this vision requires action, interaction, and memory to work together, so that each encounter can inform current decisions and contribute experience for future tasks. background 0.92 0.05 0.05 asserted 0.99
We present EvoLife, a unified navigation harness that enables a multimodal embodied agent to choose how to move, when to ask, and what experience to carry forward. contribution 0.99 0.91 0.09 asserted 1.00
Its hybrid action space supports waypoint navigation, local discrete adjustments, pixel-based navigation, and object-directed navigation. method 0.95 0.75 0.03 asserted 1.00
Active interaction allows the agent to seek clarification or assistance when it encounters ambiguity or difficulty, while persistent spatial memory binds observations and dialogue to places, retaining useful experience to guide subsequent visits. method 0.95 0.43 0.06 asserted 1.00
We evaluate eight benchmarks across four task families: vision-language navigation, object navigation, interactive instance navigation, and continuous multi-task navigation. method 0.91 0.91 0.03 asserted 0.98
With an average of only 1.87 active interactions per episode, our method achieves a first-visit success rate (SR) of 100% across four 100-episode sets spanning vision-language navigation and object navigation. result 1.00 0.96 0.14 asserted 1.00
Complete-sequence SR reaches 98% on 50 LHPR-VLN sequences and 94% on 50 GOAT-Bench sequences. result 1.00 0.94 0.06 asserted 1.00
In the R2R revisit setting, success weighted by path length (SPL) increases from 82.23% to 93.31% with retained experience. result 1.00 0.95 0.06 asserted 1.00
Multi-day physical navigation further demonstrates exploration, proactive help-seeking, and revisiting remembered places. result 0.88 0.89 0.07 asserted 1.00
Together, these results advance the vision of bringing embodied agents to life through autonomous action, proactive interaction, and the accumulation and reuse of experience. interpretation 1.00 0.23 0.85 asserted 0.97

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.80

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-30: 1 request, 8,324 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (6)
Figure 1 of the paper
Figure 1. EvoLife takes a step toward bringing embodied agents to life through autonomous action, proactive interaction, and the accumulation of experience. The left panel presents four task families: vision-language navigation [2, 3, 4], object navigation [5, 6, 7], interactive instance navigation [8, 9], and continuous multi-task navigation [10, 11]. The center shows the unified harness, connecting a hybrid action space, active interaction, and persistent spatial memory. The right panel reports success rate (SR) across eight simulation benchmarks and illustrates multi-day physical…
Figure 2 of the paper
Figure 2. Overview of the unified EvoLife framework. A multimodal agent integrates hybrid actions, active interaction, and persistent memory within a unified decision loop. The agent adaptively selects navigation actions based on current observations and execution feedback, actively seeks external guidance to resolve navigation uncertainties, and retains spatially grounded experience for future tasks. Ci denotes a geometric candidate waypoint. Pi denotes a remembered place associated with its recorded pose and supporting evidence. Di denotes the i-th tool-level decision, encompassing…
Figure 3 of the paper
Figure 3. Online guide response generation. The navigation agent actively generates a question, while an independent multimodal guide generates a natural-language response based on the navigation instruction, current observations, target information, and reference route. The figure illustrates the current and target views, reference route, and spatial relationships between the agent and the destination. Only the generated natural-language response is returned to the navigation agent, which independently makes subsequent decisions. This process simulates natural human-to-human interaction,…
Figure 4 of the paper
Figure 4. Qualitative example of active interaction and persistent memory in R2R-CE. The agent actively asks whether a nearby wooden door is the intended destination or whether it should continue through a narrow archway. Based on the guide’s response, the agent adjusts its navigation direction, proceeds through the archway, and records its entrance as place P2. Q2 and A2 denote the agent’s second interaction question and the guide’s corresponding answer, respectively, while P2 denotes the stored place record. decisions. In contrast, Qwen3.8-Max and DeepSeek V4.1-Flash adopt more conservative…
Figure 5 of the paper
Figure 5. Failure case analysis on GOAT-Bench. Left: the agent successfully completes the picture-finding subtask. Right: the agent subsequently identifies the bookcase with diamond-shaped compartments through active interaction, but stops 0.73 m from the reference target, exceeding the required 0.25 m success threshold. Orange dots on the maps indicate the agent’s position. Pi denotes a saved place in persistent memory. Ci labels candidate waypoint i. the agent actively asks whether the wooden door is the intended destination or whether it should continue through the archway. The guide rules…
Figure 6 of the paper
Figure 6. Three consecutive real-world robot navigation tasks across two days. In Task 1, the robot navigates from room 625 to the elevator on the opposite side of the floor, recording key locations along the route. In Task 2, the robot returns to room 625 using persistent memory and actively requests human assistance when encountering a closed door. In Task 3, the robot identifies and restores the matching saved map on the following morning and navigates to the glass office visited the previous day. The three trajectories are visualized in a shared coordinate frame, demonstrating active…

Figures are the authors', reproduced from the paper.

References (86)
  1. [1] A. M. Turing, “Computing Machinery and Intelligence,”Mind, vol. LIX, no. 236, pp. 433–460, 1950. 17 on Pith
  2. [2] P. Anderson, Q. Wu, D. Teney, J. Bruce, M. Johnson, N. Sünderhauf, I. Reid, S. Gouldet al., “Vision-and-Language Navigation: Interpreting Visually-Grounded Navigation Instructions in Real Environments,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 201... on Pith
  3. [3] A. Ku, P. Anderson, R. Patel, E. Ie, and J. Baldridge, “Room-Across-Room: Multilingual Vision-and-Language Navigation with Dense Spatiotemporal Grounding,” inProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020, pp. 4392–4412 on Pith
  4. [4] J. Krantz, E. Wijmans, A. Majumdar, D. Batra, and S. Lee, “Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments,” inComputer Vision – ECCV 2020, 2020, pp. 104–120 on Pith
  5. [5] S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. Turner, E. Undersander, W. Galuba et al., “Habitat-Matterport 3D Dataset (HM3D): 1000 Large-scale 3D Environments for Embodied AI,” inAdvances in Neural Information Processing Systems Datasets and Benchmar... on Pith
  6. [6] K. Yadav, R. Ramrakhya, S. K. Ramakrishnan, T. Gervet, J. Turner, A. Gokaslan, N. Maestre, A. X. Chang et al., “Habitat-Matterport 3D Semantics Dataset,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 4927–4936 on Pith
  7. [7] N. Yokoyama, R. Ramrakhya, A. Das, D. Batra, and S. Ha, “HM3D-OVON: A Dataset and Benchmark for Open-Vocabulary Object Goal Navigation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 5543–5550 on Pith
  8. [8] W. Huang, S. Zhu, M. Wei, S. Zhang, J. Xu, X. Liu, H. Wang, T. Wanget al., “VL-LN Bench: Towards Long-horizon Goal-oriented Navigation with Active Dialogs,” arXiv preprint arXiv:2512.22342, 2025 on PithPith's reviewarXiv
  9. [9] F. Taioli, E. Zorzi, G. Franchi, A. Castellini, A. Farinelli, M. Cristani, and Y. Wang, “Collaborative Instance Object Navigation: Leveraging Uncertainty-Awareness to Minimize Human-Agent Dialogues,” in2025 IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp.... on Pith
  10. [10] X. Song, W. Chen, Y. Liu, W. Chen, G. Li, and L. Lin, “Towards Long-Horizon Vision-Language Navigation: Platform, Benchmark and Method,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 12078–12088 on Pith
  11. [11] M. Khanna, R. Ramrakhya, G. Chhablani, S. Yenamandra, T. Gervet, M. Chang, Z. Kira, D. S. Chaplotet al., “GOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16373–16383 on Pith
  12. [12] H. Dai, X. Wang, L. Wang, K. Sheng, Z. He, C. Liu, W. Ye, and Q. Chen, “Dped-vln: A benchmark for socially compliant vision-and-language navigation in dynamic pedestrian environments,”arXiv preprint arXiv:2609.21504, 2026 on PithPith's reviewarXiv
  13. [13] J. Thomason, M. Murray, M. Cakmak, and L. Zettlemoyer, “Vision-and-Dialog Navigation,” inProceedings of the Conference on Robot Learning, ser. Proceedings of Machine Learning Research, vol. 100, 2020, pp. 394–406 on Pith
  14. [14] K. Nguyen and H. Daumé III, “Help, Anna! Visual Navigation with Natural Multimodal Assistance via Retro- spective Curiosity-Encouraging Imitation Learning,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint... on Pith
  15. [15] Y. Dai, R. Peng, S. Li, and J. Chai, “Think, Act, and Ask: Open-World Interactive Personalized Robot Navigation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 3296–3303 on Pith
  16. [16] J. Krantz, S. Banerjee, W. Zhu, J. Corso, P. Anderson, S. Lee, and J. Thomason, “Iterative Vision-and-Language Navigation,” in2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 14921–14930 on Pith
  17. [17] L. Wang, K. Sheng, Z. He, J. Li, Y. Qin, H. Dai, X. Wang, J. Yanget al., “A Comprehensive Survey and Systematic Real-World Evaluation of Embodied Vision-and-Language Navigation,”IEEE Transactions on Automation Science and Engineering, vol. 23, pp. 14405–14434, 2026 on Pith
  18. [18] L. Wang, X. Xia, H. Zhao, H. Wang, T. Wang, Y. Chen, C. Liu, Q. Chenet al., “Rethinking the embodied gap in vision-and-language navigation: A holistic study of physical and visual disparities,” in2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025, pp.... on Pith
  19. [19] D. An, H. Wang, W. Wang, Z. Wang, Y. Huang, K. He, and L. Wang, “ETPNav: Evolving Topological Planning for Vision-Language Navigation in Continuous Environments,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 7, pp. 5130–5145, 2025 on Pith
  20. [20] G. Zhou, Y. Hong, and Q. Wu, “NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 7, 2024, pp. 7641–7649 on Pith
show 66 more
  1. [21] G. Zhou, Y. Hong, J. Zhang, X. Zhao, J. Zhou, Z. Lei, Z. Wang, C. Zhaoet al., “NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime,” arXiv preprint arXiv:2610.10787, 2026. 18 on PithPith's reviewarXiv
  2. [22] Y. Xu, Y. Pan, and Z. Liu, “Dream to Recall: Imagination-Guided Experience Retrieval for Memory-Persistent Vision-and-Language Navigation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 8, pp. 9035–9049, 2026 on Pith
  3. [23] X. Zhao, J. Zhou, S. Lin, G. Zhou, Z. Li, X. Yan, J. Liu, A. van den Hengelet al., “NavHarness: Towards Lifelong Embodied Navigation,” arXiv preprint arXiv:2609.34276, 2026 on PithPith's reviewarXiv
  4. [24] Y. Chen, L. Che, Z. Huang, W. Fu, C. Wang, X. Cao, D. Liu, Y. Yanget al., “HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness,” arXiv preprint arXiv:2609.15195, 2026 on PithPith's reviewarXiv
  5. [25] L. Wang, Z. He, R. Dang, M. Shen, C. Liu, and Q. Chen, “Vision-and-language navigation via causal learning,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 13139–13150 on Pith
  6. [26] L. Wang, Z. He, M. Shen, J. Yang, C. Liu, and Q. Chen, “Magic: Meta-ability guided interactive chain-of- distillation for effective-and-efficient vision-and-language navigation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026 on Pith
  7. [27] J. Zhang, K. Wang, R. Xu, G. Zhou, Y. Hong, X. Fang, Q. Wu, Z. Zhanget al., “NaVid: Video-based VLM Plans the Next Step for Vision-and-Language Navigation,” inRobotics: Science and Systems XX, 2024 on Pith
  8. [28] M. Wei, C. Wan, X. Yu, T. Wang, Y. Yang, X. Mao, C. Zhu, W. Caiet al., “StreamVLN: Streaming Vision-and- Language Navigation via SlowFast Context Modeling,” in2026 IEEE International Conference on Robotics and Automation (ICRA), 2026, pp. 22174–22181 on Pith
  9. [29] A.-C. Cheng, Y. Ji, Z. Yang, Z. Gongye, X. Zou, J. Kautz, E. Bıyık, H. Yinet al., “NaVILA: Legged Robot Vision-Language-Action Model for Navigation,” inRobotics: Science and Systems XXI, 2025 on Pith
  10. [30] K. Sheng, L. Wang, H. Dai, J. Li, Y. Qin, Z. He, C. Liu, and Q. Chen, “P2DNav: Panorama-to-Downview Reasoning for Zero-shot Vision-and-Language Navigation,” arXiv preprint arXiv:2605.19634, 2026 on PithPith's reviewarXiv
  11. [31] K. Sheng, L. Wang, J. Li, H. Dai, C. Liu, and Q. Chen, “NavJev: Efficient Vision-Language Navigation via Action- Centric Visual Compression and Discriminative Action-Semantic Memory,” arXiv preprint arXiv:2609.34969, 2026 on PithPith's reviewarXiv
  12. [32] L. Wang, Z. He, J. Li, R. Xia, M. Hu, C. Yao, C. Liu, Y. Tanget al., “CLASH: Collaborative Large-Small Hierarchical Framework for Continuous Vision-and-Language Navigation,” arXiv preprint arXiv:2512.10360, 2025 on PitharXiv
  13. [33] B. Lin, Y. Nie, Z. Wei, Y. Zhu, H. Xu, S. Ma, J. Liu, and X. Liang, “Correctable Landmark Discovery via Large Models for Vision-Language Navigation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 8534–8548, 2024 on Pith
  14. [34] B. Lin, Y. Nie, Z. Wei, J. Chen, S. Ma, J. Han, H. Xu, X. Changet al., “NavCoT: Boosting LLM-Based Vision-and-Language Navigation via Learning Disentangled Reasoning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 7, pp. 5945–5957, 2025 on Pith
  15. [35] Z. Yu, Y. Long, Z. Yang, C. Zeng, H. Fan, J. Zhang, and H. Dong, “Correctnav: Self-correction flywheel empowers vision-language-action navigation model,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 22, 2026, pp. 18737–18745 on Pith
  16. [36] S. Zhu, W. Huang, S. Zhang, Y. Yang, M. Wei, C. Wan, T. Liu, T. Wanget al., “GG-Nav: Goal-oriented Grounding Chain-of-Thought Elicits Reasoning in Navigation Foundation Models,” arXiv preprint arXiv:2610.14541, 2026 on PithPith's reviewarXiv
  17. [37] J. Kwon, S. Lee, H. Park, K. Min, and J. Ok, “ProCompNav: Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries,” inAdvances in Neural Information Processing Systems (NeurIPS), 2026, accepted, oral presentation on Pith
  18. [38] Z. Li, S. Lin, Y. Shao, J. Zhang, X. Shi, S. Li, and Q. Wu, “Talk2Escape: Conversational Grounding for Vision-and-Language Navigation,” in2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026 on Pith
  19. [39] M. Liu, J. Wang, Y. Huang, H. Su, J. Lv, Y. Ma, and Y. Liu, “AVERT-VLN: Abstention-aware Visual Error Recovery and Training for Vision-and-Language Navigation,” arXiv preprint arXiv:2609.39579, 2026 on PithPith's reviewarXiv
  20. [40] C. Huang, O. Mees, A. Zeng, and W. Burgard, “Visual Language Maps for Robot Navigation,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 10608–10615 on Pith
  21. [41] Q. Gu, A. Kuwajerwala, S. Morin, K. M. Jatavallabhula, B. Sen, A. Agarwal, C. Rivera, W. Paulet al., “ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and Planning,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 5021–5028 on Pith
  22. [42] Y. Yang, H. Yang, J. Zhou, P. Chen, H. Zhang, Y. Du, and C. Gan, “3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 17294–17303 on Pith
  23. [43] M. Savva, A. Kadian, O. Maksymets, Y. Zhao, E. Wijmans, B. Jain, J. Straub, J. Liuet al., “Habitat: A Platform for Embodied AI Research,” in2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 9338–9346. 19 on Pith
  24. [44] OpenAI, “GPT-6 Astra System Card,” Sep. 2026 on Pith
  25. [45] Qwen Team, “Qwen3.8-Max: A new bar for coding and cowork,” Aug. 2026 on Pith
  26. [46] DeepSeek-AI, “DeepSeek-V4.1-Flash: Pushing the limits of KV cache compression,” 2026 on Pith
  27. [47] P. Anderson, A. Chang, D. S. Chaplot, A. Dosovitskiy, S. Gupta, V. Koltun, J. Kosecka, J. Maliket al., “On Evaluation of Embodied Navigation Agents,” arXiv preprint arXiv:1807.06757, 2018 on PithPith's reviewarXiv
  28. [48] G. Ilharco, V. Jain, A. Ku, E. Ie, and J. Baldridge, “General Evaluation for Instruction Conditioned Navigation using Dynamic Time Warping,” inVisually Grounded Interaction and Language (ViGIL), NeurIPS 2019 Workshop, 2019 on Pith
  29. [49] Z. He, L. Wang, S. Li, Q. Yan, C. Liu, and Q. Chen, “A Multilevel Attention Network with Sub-Instructions for Continuous Vision-and-Language Navigation,”Applied Intelligence, vol. 55, no. 7, p. 657, 2025 on Pith
  30. [50] J. Zhang, K. Wang, S. Wang, M. Li, H. Liu, S. Wei, Z. Wang, Z. Zhanget al., “Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks,” inRobotics: Science and Systems XXI, 2025 on Pith
  31. [51] Z. Zhang, W. Zhu, H. Pan, X. Wang, R. Xu, X. Sun, and F. Zheng, “ActiveVLN: Towards Active Exploration via Multi-Turn RL in Vision-and-Language Navigation,” in2026 IEEE International Conference on Robotics and Automation (ICRA), 2026, pp. 12611–12618 on Pith
  32. [52] Z. Xin, W. Li, Y. Jiang, B. Wang, R. Cong, J. Qin, and S. Huang, “DecoVLN: Decoupling Observation, Reasoning, and Correction for Vision-and-Language Navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2026, pp. 32410–32420 on Pith
  33. [53] Q. Wang, M. Li, J. Guan, J. Ye, S. Xie, Y. Liu, J. Chen, Z. Lianget al., “Qwen-VLA: Unifying Vision-Language- Action Modeling across Tasks, Environments, and Robot Embodiments,” arXiv preprint arXiv:2605.30280, 2026 on PithPith's reviewarXiv
  34. [54] Intern Robotics, “InternVLA-N1: An Open Dual-System Vision-Language Navigation Foundation Model with Learned Latent Plans,” 2025, technical report, Shanghai AI Laboratory on Pith
  35. [55] R. Dang, J. Guo, B. Hou, S. Leng, K. Li, X. Li, J. Liu, Y. Maoet al., “RynnBrain: Open Embodied Foundation Models,” arXiv preprint arXiv:2602.14979, 2026 on PitharXiv
  36. [56] S. Zeng, D. Qi, X. Chang, F. Xiong, S. Xie, X. Wu, S. Liang, M. Xuet al., “JanusVLN: Decoupling Semantics and Spatiality with Dual Implicit Memory for Vision-Language Navigation,” inInternational Conference on Learning Representations (ICLR), 2026 on Pith
  37. [57] M. Wei, C. Wan, J. Peng, X. Yu, Y. Yang, D. Feng, W. Cai, C. Zhuet al., “Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and-Language Navigation,” inInternational Conference on Learning Representations (ICLR), 2026 on Pith
  38. [58] J. Chen, J. Hu, H. Bai, M. Luo, X. Xue, B. Ren, C. Bai, S. Xieet al., “AstraNav-World: World Model for Foresight Control and Consistency,” arXiv preprint arXiv:2512.21714, 2025 on PithPith's reviewarXiv
  39. [59] S. Wang, A. Luo, F. Huang, J. Xu, X. Wang, Y. Wang, Q. Ma, F. Yanget al., “LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation,” 2026 on Pith
  40. [60] A.-C. Cheng, I. Liu, E. Bu, J. Bjorck, H. Yin, Z. Luo, J. Kautz, L. J. Fanet al., “ASENA: Self-evolving Agents for Embodied Navigation,” arXiv preprint arXiv:2609.39207, 2026 on PithPith's reviewarXiv
  41. [61] X. Xue, J. Hu, M. Luo, S. Xie, J. Chen, Z. Xie, K. Quan, W. Guoet al., “OmniNav: A Unified Frame- work for Prospective Exploration and Visual-Language Navigation,” inInternational Conference on Learning Representations (ICLR), 2026 on Pith
  42. [62] J. Zhang, G. Zhou, H. Yin, Y. Huang, Z. Lei, Q. Peng, H. Yuan, J. Zhanget al., “Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System,” arXiv preprint arXiv:2606.18112, 2026 on PithPith's reviewarXiv
  43. [63] A. Bounhar, A. Somani, A. Kabra, A. Valente, A. Petralia, A. Sade, A. Jeffares, A. Jianget al., “Robostral Navigate,” arXiv preprint arXiv:2607.20785, 2026 on PithPith's reviewarXiv
  44. [64] N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher, “VLFM: Vision-Language Frontier Maps for Zero-Shot Semantic Navigation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 42–48 on Pith
  45. [65] H. Yin, X. Xu, Z. Wu, J. Zhou, and J. Lu, “SG-Nav: Online 3D Scene Graph Prompting for LLM-based Zero-shot Object Navigation,” inAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, Eds., vol. 37.... on Pith
  46. [66] Z. Zhu, X. Wang, Y. Li, Z. Zhang, X. Ma, Y. Chen, B. Jia, W. Lianget al., “Move to Understand a 3D Scene: Bridging Visual Grounding and Exploration for Efficient and Versatile Embodied Navigation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV... on Pith
  47. [67] J. Zhang, A. Li, Y. Qi, M. Li, J. Liu, S. Wang, H. Liu, G. Zhouet al., “Embodied Navigation Foundation Model,” inInternational Conference on Learning Representations (ICLR), C. Vondrick, B. Hariharan, C. Raffel, L. Pinto, D. Yang, and A. Faust, Eds., vol. 2026, 2026, pp. 127293–127322 on Pith
  48. [68] Z. Chu, S. Xie, X. Wu, Y. Shen, M. Luo, Z. Wang, F. Liu, X. Lenget al., “ABot-N0: Technical Report on the VLA Foundation Model for Versatile Embodied Navigation,” arXiv preprint arXiv:2602.11598, 2026 on PitharXiv
  49. [69] M. Zhang, Y. Du, C. Wu, J. Zhou, Z. Qi, J. Ma, and B. Zhou, “ApexNAV: An Adaptive Exploration Strategy for Zero-Shot Object Navigation With Target-Centric Semantic Fusion,”IEEE Robotics and Automation Letters, vol. 10, no. 11, pp. 11530–11537, 2025 on Pith
  50. [70] J. Zhang, J. Xu, Y. Yu, J. Guo, R. Hu, H. Bao, X. Zhou, and S. Peng, “SuperNav: An Agentic Navigation System for Any Task in Any Scene,” arXiv preprint arXiv:2610.12126, 2026 on PithPith's reviewarXiv
  51. [71] G. Liu, S. Ling, R. Liu, Z. Gong, and J. Hu, “AECNav: Active Evidence Consolidation for Efficient Zero-Shot Open-Vocabulary Object Navigation,” arXiv preprint arXiv:2608.10817, 2026 on PithPith's reviewarXiv
  52. [72] D. Zheng, S. Huang, L. Zhao, Y. Zhong, and L. Wang, “Towards Learning a Generalist Model for Embodied Navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 13624–13634 on Pith
  53. [73] S. Wang, Y. Wang, W. Li, X. Cai, Y. Wang, M. Chen, K. Wang, Z. Suet al., “Aux-Think: Exploring Reasoning Strategies for Data-Efficient Vision-Language Navigation,” inAdvances in Neural Information Processing Systems, D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Gh... on Pith
  54. [74] J. Zuo, L. Mu, F. Jiang, C. Ma, M. Xu, and Y. Qi, “FantasyVLN: Unified Multimodal Chain-of-Thought Reasoning for Vision-and-Language Navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 29674–29683 on Pith
  55. [75] S. Wang, B. Liu, Z. Gao, L. Ma, X. Wang, Y. Xie, and X. Tan, “Explore with Long-term Memory: A Benchmark and Multimodal LLM-based Reinforcement Learning Framework for Embodied Exploration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CV... on Pith
  56. [76] H. Niu, L. Zhang, X. Ji, R. Ying, P. Liu, and F. Wen, “SSMG-Nav: Enhancing Lifelong Object Navigation with Semantic Skeleton Memory Graph,” in2026 IEEE International Conference on Robotics and Automation (ICRA), 2026, pp. 2258–2265 on Pith
  57. [77] Z. Ge, X. Jia, C. Wu, Y. Zhou, Z. Wu, and Y.-G. Jiang, “EvoMemNav: Efficient Self-Evolving Fine-Grained Memory for Zero-Shot Embodied Navigation,” arXiv preprint arXiv:2606.03509, 2026 on PithPith's reviewarXiv
  58. [78] G. Zhang, M. Ding, J. Wu, R. Liao, and V. Tresp, “ReEXplore: Improving MLLMs for Embodied Exploration with Contextualized Retrospective Experience Replay,”arXiv preprint arXiv:2511.19033, 2025 on PitharXiv
  59. [79] J. Hu, X. Xue, B. Ren, M. Luo, J. Chen, H. Bai, L. You, and M. Xu, “AstraNav-Memory: Contexts Compression for Long Memory,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 8097–8109 on Pith
  60. [80] Y. Chen, Z. Huang, W. Fu, D. Peng, S.-Y. Tian, K.-Y. Yu, and L.-Z. Guo, “STEGNav: Spatio-Temporal Event Graph Reasoning for Multimodal Lifelong Object Navigation,” arXiv preprint arXiv:2608.28279, 2026 on PithPith's reviewarXiv
  61. [81] Y. Lu, Y. Du, D. Liu, Y. Zhou, C. Wang, and Y. Yin, “GSMem: 3D Gaussian Splatting as Persistent Spatial Memory for Zero-Shot Embodied Exploration and Reasoning,” arXiv preprint arXiv:2603.19137, 2026 on PitharXiv
  62. [82] X. Li, F. Lyu, H. Wu, M. Liu, J.-N. Liu, and G. Liu, “Stop Wandering: Efficient Vision-Language Navigation via Metacognitive Reasoning,” arXiv preprint arXiv:2604.02318, 2026 on PitharXiv
  63. [83] P. Chen, G. Zhang, J. Ma, and Q. Li, “Hypothesis Graph Refinement: Hypothesis-Driven Exploration with Cascade Error Correction for Embodied Navigation,” inComputer Vision – ECCV 2026. Cham: Springer Nature Switzerland, 2026, pp. 556–573 on Pith
  64. [84] T. Lee, Y. Jang, J. Heo, J. Choi, and H. J. Kim, “ObsGraph: Hierarchical Observation Representation for Embodied Reasoning and Exploration,” arXiv preprint arXiv:2606.24068, 2026 on PithPith's reviewarXiv
  65. [85] J. Li, B. Wang, J. Xia, M. Li, and S. Hu, “HIMM: Human-Inspired Long-Term Memory Modeling for Embodied Exploration and Question Answering,” 2026 on Pith
  66. [86] A. Schperberg, S. K. Panda, A. P. Vinod, B. Rokaha, M. K. Jawed, and S. Di Cairano, “RoboAtlas: Contextual Active SLAM,” arXiv preprint arXiv:2606.26046, 2026. 21 A Additional Qualitative Examples We present one example from each of the eight benchmarks to examine howEvoLifeco... on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment