Pith. sign in

REVIEW 3 major objections 4 minor 49 references

Handing a general-purpose agent nothing but a single RGB camera and four discrete actions yields navigation success on par with trained, large-scale policies.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 00:37 UTC pith:RGCY2RZ2

load-bearing objection The minimal-interface zero-shot VLN numbers are real and well reported, but the 'rival industrial-scale policies' headline is one same-split trained baseline short; solid paper that deserves a review with revisions. the 3 major comments →

arxiv 2607.26148 v1 pith:RGCY2RZ2 submitted 2026-07-28 cs.RO

Embodied Agents Take Control: Minimal-Interface Zero-Shot Agents Rival Industrial-Scale Policies in Vision-and-Language Navigation

Jian Zhou , Xunyi Zhao , Gengze Zhou , Zerui Li , Sihao Lin , Jiajun Liu , Qi Wu This is my paper
classification cs.RO
keywords embodied agentsvision-and-language navigationzero-shot navigationagentic controlminimal interfacehybrid interfaceslong-horizon navigationvision-language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that agentic embodied control—where a general-purpose model, not a task-specific policy or workflow, directs the interaction loop—already works in zero-shot navigation. With only a front-facing monocular frame and four primitive actions, unmodified software-engineering agent harnesses reach 70.7±3.5% success at default reasoning effort and 78% at maximum effort on the standard zero-shot continuous-navigation test set. The authors argue this is competitive with engineered zero-shot systems and recent industrial-scale trained navigators, despite receiving no map, depth, pose, memory, or navigation training. They also show that capability is largely model-centered: model choice dominates, harness differences are descriptive, and a trained waypoint interface helps weak models but can constrain strong ones—until it is exposed as an optional tool, which makes the strongest agent better, faster, and cheaper.

Core claim

The paper's central claim is that a minimal-interface agent—one monocular RGB camera, four discrete actions, and a generic software-engineering harness—can hold the entire embodied decision loop and navigate as well as systems with maps, depth, waypoints, explicit memory, and navigation-specific training. Replicated default-effort runs average 70.7±3.5% success (and a strong configuration reaches 78% at maximum reasoning effort) on the standard zero-shot R2R-CE test set. When a trained waypoint predictor is exposed alongside primitives as an optional capability rather than a forced control path, a default-effort agent reaches 76.7±0.6% success while using half the environment steps and less

What carries the argument

The minimal-interface probe: two tools exposed to a frozen general-purpose model—observe(), returning a single 512×512 egocentric RGB frame, and step(), executing an ordered list of four discrete primitives (forward 0.25 m, turn left/right 15°, stop). The model, not a policy or a workflow, decides when to look, what to do, how to recover, and when to stop; the harness merely maintains the session and executes tool calls. In the hybrid study, a trained depth-based waypoint predictor is added as an optional tool, letting the agent compose coarse waypoint moves with fine primitive control and decide per step whether the trained module is worth invoking. This setup isolates the model–harness–int

Load-bearing premise

The headline claim that minimal-interface agents rival trained policies assumes the fixed 100-episode rand100 subset is a fair proxy for the full validation-unseen benchmark; the paper compares zero-shot results on this subset against trained policies on the full split, and the paper itself notes that trained rows serve only as a reference.

What would settle it

Run the same minimal-interface agents on the full validation-unseen split (or a fresh random sample of comparable size) and compare success rates against trained policies evaluated on the same split. If the agent's success rate falls well below the reported 70–78% range—or if the rand100 subset turns out to have substantially shorter routes than the full set—the central claim of competitiveness collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Navigation capability can be obtained without task-specific training, maps, depth, pose, or explicit memory, as long as a capable general-purpose model holds the interaction loop.
  • Imposing a trained interface (e.g., waypoints) helps weak primitive controllers but can hurt strong ones; exposing the same module as an optional tool is more effective and more efficient.
  • Model choice is the dominant factor: on the fixed minimal interface, success rate spans from roughly 5% to 72% as only the model changes, while reasoning-effort changes are model-specific and mostly small.
  • Longer-horizon routes remain a hard limit: success falls to 26–39%, interaction context grows without bound, and failures are silent—most failed episodes end with a confident, incorrect claim of success.
  • The same minimal interface transfers to a physical quadruped, where reasoning and perception carry over but body awareness, calibrated motion, and spatial memory do not.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The headline comparison to trained policies rests on a fixed 100-episode subset of the zero-shot benchmark, while trained rows are reported on the full validation split; a same-split evaluation would be needed to confirm the 'rival industrial-scale' framing.
  • The model-centered result suggests that embodiment-specific post-training of the same general-purpose agents could push success further and may reduce the need for scaffolding altogether.
  • The hybrid result generalizes as a design principle: rather than hard-coding a module into the control path, expose it as an optional tool and let the agent decide when to invoke it—likely applicable beyond navigation.
  • The silent-failure analysis implies that practical deployment of such agents requires an external verification layer (e.g., a map or stop-validation) because the model's own confidence is uncorrelated with actual success.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces 'agentic embodied control,' in which unmodified software-engineering agent harnesses control a robot directly with one monocular RGB view and four discrete actions, with no navigation-specific policy, map, memory, or workflow. On the fixed rand100 subset of R2R-CE val-unseen, replicated default-effort runs reach 70.7±3.5% SR (opus-5) and 68.3±1.5% SR (fable-5), while a single max-effort run reaches 78% SR. Adding a trained waypoint module as an optional tool raises default-effort fable-5 to 76.7±0.6% SR with lower step count and wall time. The paper ablates model, harness, and interface; analyzes all 25 failures in one run in detail; tests longer-horizon RxR-CE episodes; and reports qualitative deployment on a physical robot.

Significance. If the evaluation is sound, the contribution is significant: the paper challenges the assumption that strong VLN-CE performance requires navigation-specific scaffolding, while also identifying model-, harness-, and interface-level bottlenecks. The strengths are real: headline default-effort numbers are replication means, the failure taxonomy is a frame-by-frame audit of all failures in a run, the ablations vary one axis at a time for most cells, and the physical-robot study is honest about its limitations. The main risk is evaluation fairness of the 'rival industrial-scale policies' claim, which rests on comparing the authors' rand100 subset against trained policies reported on the full val-unseen split. The manuscript discloses this mismatch, but the abstract, title, and figure framing nevertheless make the cross-split comparison the headline.

major comments (3)
  1. [§4.1, Table 1, §4.2] The headline comparison of minimal zero-shot rows (70.7±3.5, 78) with industrial-scale trained policies (Qwen-RobotNav 72 SR, ABot-N1 71 SR) is cross-split: zero-shot rows use the fixed rand100 subset of val-unseen, while trained rows use the full val-unseen split. The manuscript discloses this ('Trained rows use the full val-unseen split, so they serve only as a reference') and calls the comparison 'not controlled,' but the title, abstract, and Fig. 1 frame the result as 'rivaling' those policies. Since rand100 is inherited from the authors' own Open-Nav work and no same-split trained baseline is provided, the comparative claim is not established. Please add at least one trained baseline evaluated on the same 100 episodes, or evaluate the minimal agents on the full val-unseen split; at minimum, temper all 'rival'/'competitive in context' claims to remain valid for the subset until this
  2. [§4.5, Table 6; §4.6] There is an inconsistency in the fable-5/Claude SDK default-effort numbers. Table 2 reports 68.3±1.5 SR over three replications on the rand100 subset, but Table 6 lists R2R-CE primitives as 75 SR under what is described as the same frozen configuration, and §4.6 begins by saying 'one complete fable-5–Claude SDK default-effort run (SR 75).' Unless Table 6 uses a different 100-episode random sample, these numbers cannot both be correct; if a different sample is used, the text must say so explicitly. This matters because the failure taxonomy in §4.6 is presented as a board run, yet its SR 75 matches Table 6, not Table 2's replication mean.
  3. [Table 4, §4.3] The interface ablation claims to hold the model, harness, and episodes fixed within each pair, but the footnote for the GPT rows shows otherwise: the gpt-5.5 pair changes the Codex CLI version (0.142.0 vs 0.145.0), and the gpt-5.6-sol pair changes both effort (low vs medium) and version. This violates the single-axis design for two of the eight rows. The Qwen and Claude pairs are controlled, and the general conclusion likely survives without the GPT rows, but the text should either remove the GPT rows from the 'holding ... fixed' claim or provide matched runs.
minor comments (4)
  1. [Abstract, §4.1] The abstract says 'standard zero-shot R2R-CE test set'; the evaluated set is actually the fixed rand100 subset of R2R-CE val-unseen. Please make this explicit in the abstract, since it is central to interpreting the headline numbers.
  2. [Table 2, Abstract] The 78% max-effort result is a single un-replicated run. It appears in the abstract as a headline. Please mark it clearly as a single run in the abstract and text, not only in Table 2's note.
  3. [Fig. 1] The right panel visually aligns the probe bar with Qwen-RobotNav's full-val-unseen bar. The text explains the split mismatch, but the figure would be clearer with a visual separator or an explicit 'different split' annotation.
  4. [§4.5] The latency statement 'Across 4,700 logged simulator episodes, median wall time is 206 s' should state which configurations are included. As written, it is easy to misread as a claim about the minimal cells only.

Circularity Check

0 steps flagged

No circular derivation: the headline SR numbers are direct measurements on a fixed external subset; the remaining concern is a split-mismatched comparison, not a circular reduction.

full rationale

The paper's central results (70.7±3.5 SR, 78 SR, 76.7±0.6 SR) are direct measurements of Habitat-native success on a fixed 100-episode subset of R2R-CE val-unseen. No parameter is fitted to the success metric, no output is defined in terms of an input, and the ablations are empirical contrasts rather than derived predictions. The headline comparison to trained industrial systems uses a different evaluation split, and the paper explicitly flags this: 'Trained rows use the full val-unseen split, so they serve only as a reference for the performance range' (§4.2) and 'the trained results use a different split' (Introduction). That is an evaluation-fairness limitation, not an equivalence by construction. The rand100 subset and the waypoint predictor originate in the authors' own prior work (Open-Nav [12], SmartWay [13], VLNVerse [38]), and some comparison systems share lab lineage; this is a provenance/self-selection concern that lowers confidence in how far the results generalize, but the measured SR values are externally reproducible and do not reduce to those citations. The hybrid result is measured, not predicted, and the paper does not claim to test the Section 5 bootstrapping cycle ('We do not test this cycle here'). Overall, the derivation chain is empirical and self-contained against an external benchmark; the minor self-citation provenance is not load-bearing in the sense required for circularity, so no significant circularity is found.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper introduces no fitted constants; its protocol constants are fixed. It relies on several unverified domain assumptions: representativeness of the n=100 subset, validity of the VLN-CE success metric as a competence measure, true general-purpose provenance of the harnesses, comparability of vendor effort/model labels, and no leakage of privileged simulator state. No new entities are postulated.

axioms (5)
  • domain assumption The fixed rand100 subset of R2R-CE val-unseen (n=100) is representative enough to support claims about zero-shot navigation performance.
    All headline numbers are measured on 100 episodes selected by prior Open-Nav work; many cells have a single run, so generalization to the full benchmark is assumed. Section 4.1.
  • domain assumption The standard VLN-CE success criterion (agent-issued STOP within 3 m geodesic of the goal) captures the intended notion of navigation capability.
    The paper adopts this metric without justifying it as a proxy for embodied agent competence; the failure taxonomy shows the metric can be gamed by object-anchored stopping. Sections A.1 and B.
  • domain assumption The three software-engineering harnesses are effectively general-purpose and contain no navigation-specific control logic.
    The claim that capability is 'model-centered' depends on the harnesses not smuggling navigation logic; the authors assert this by provenance, not by analysis. Section 3.
  • domain assumption Vendor reasoning-effort labels and model identities are comparable enough that model-choice differences reflect model capability rather than serving-path artifacts.
    The model sweep spans 5–72 SR with one run per cell and unmatched serving paths; the authors acknowledge the harness comparison is descriptive, but the model comparison also inherits this. Sections 4.3 and A.4.
  • domain assumption The driver actually discards shortest-path/oracle sensors, so the agent receives no privileged state beyond the declared interface.
    The minimal-interface claim would collapse if privileged signals reached the model; the paper states the sensors are discarded at the tool bridge, but no verification artifact is provided. Section A.1.

pith-pipeline@v1.3.0-alltime-deepseek · 29201 in / 15446 out tokens · 149192 ms · 2026-08-01T00:37:48.862795+00:00 · methodology

0 comments
read the original abstract

Autonomous embodied agents must sustain a long decision-making loop that involves perceiving, acting, verifying, and self-correcting over many steps. Current systems sustain this loop through task-specific workflows or embodied policies. We study a third form, agentic embodied control, in which a general-purpose agent holds the loop itself. Using zero-shot navigation as a controlled testbed, we evaluate three software-engineering agent harnesses given only a monocular RGB camera and discrete actions. Under this strictly minimal condition, replicated default-effort configurations reach 70.7$\pm$3.5% success (opus-5, mean over three runs), and fable-5 reaches 78% at maximum effort. When a trained waypoint tool is exposed alongside primitives as an optional capability, the hybrid fable-5 agent reaches 76.7$\pm$0.6% at default effort, using half the environment steps and less than one quarter of the wall time of the maximum-effort primitive run. Controlled interventions show that capability is primarily model-centered: model choice strongly changes success, harness effects are descriptive, and a forced waypoint interface helps weaker models but can hinder stronger ones. Performance nevertheless falls sharply on longer-horizon tasks, while latency and context growth limit sustained operation. These results show that agentic control is already competitive in zero-shot navigation and that models, harnesses, and interfaces offer complementary paths toward autonomous embodied agents.

Figures

Figures reproduced from arXiv: 2607.26148 by Gengze Zhou, Jiajun Liu, Jian Zhou, Qi Wu, Sihao Lin, Xunyi Zhao, Zerui Li.

Figure 1
Figure 1. Figure 1: Who holds the interaction loop, what we hand the model, and how far that gets it. Left: at inference the loop is directed by an external environment loop (policy), human-authored pipeline code (workflow), or the model itself (agentic), with systems placed by the loop they execute, trained (T) or zero-shot (ZS). Middle: our minimal-interface probe, a software-engineering harness with one monocular RGB view … view at source ↗
Figure 2
Figure 2. Figure 2: One successful R2R-CE episode, reconstructed from its raw log. Six of 20 archived observations, each with the step() call issued from that view and the recorded reasoning verbatim ([. . . ] = elision). The bottom tape lists all 111 primitives in order (F = 0.25 m forward, L/R = 15◦ turn). Amber marks segments the model itself describes as blocked or off-route. No map, pose, depth, panorama, or waypoint is … view at source ↗
Figure 3
Figure 3. Figure 3: Locating capability across model, harness, and interface (shared SR scale). (A) With harness and interface fixed, the model spans 5–72 SR. (B) The open and vendor harness differ by 2–7 SR across the six shared pairs—mini-swe-agent leads in five, the vendor harness in one (descriptive; serving paths differ). (C) A waypoint interface rescues the weakest models on R2R-CE but reverses for strong models on VLNV… view at source ↗
Figure 4
Figure 4. Figure 4: ep2, category A (wrong referent / wrong branch). The kitchen has two sinks. Obs 1, 3 and 6: the route itself is right. The agent turns off the spawn wall, skirts the dining table, and enters the kitchen aisle. Obs 9 (amber): the annotated far sink is visible through the doorway at the end of the aisle, and the agent even resolves to “see if there’s another sink ahead”, a check it never executes. Obs 12: in… view at source ↗
Figure 5
Figure 5. Figure 5: ep91, category B (stop-decision failure). The cleanest exhibit of object-anchored stopping. Obs 1: spawn facing a wall. Obs 3: after turning, the geometry is parsed correctly: pool right, bar ahead-left. Obs 5: “roughly 4–5 meters away”, and it is. The estimate chain stays arithmetically consistent to the end. Obs 8: STOP with the bar corner ∼2 m away and 467 primitives unused. The failure is entirely in t… view at source ↗
Figure 6
Figure 6. Figure 6: ep59, category C (runaway search). No exercise room is ever found, so one is invented. Obs 1 and 6: the exit from the kitchen into the tiled hall is correct. The agent then clears the wrong wing room by room. Obs 56: the surrogate: a home-gym rack standing in the master bedroom, the bed visible in the same frame. Obs 62 (amber): wedged in a curtain gap (the next frame is byte-identical). Obs 96: it walks b… view at source ↗
Figure 7
Figure 7. Figure 7: ep98, category D (geometry / simulator trap). What the missing collision signal costs. Obs 12 and 26 (amber): the hallway chokepoint, the same view 67 env steps apart. ∼100 primitives spent with zero progress, every one reported as executed. Obs 41: the re-route succeeds (the dining room is reached, the kitchen doorway visible beyond the table), but chair meshes seal every path to it. Obs 55 and 72 (amber)… view at source ↗
Figure 8
Figure 8. Figure 8: Condition and reasoning navigation, success. go2 ep16. The instruction hides a logic test. The agent turns around, evaluates that the stated arithmetic is false, and drives to the yellow tape rather than the purple decoy. The branch is settled before the first step. INTERACT multi-person interaction nav · go2 ep25 “Go to the remote control that fell on the floor behind you. Assuming you could pick it up, b… view at source ↗
Figure 9
Figure 9. Figure 9: Multi person interaction, success. go2 ep25. Told to bring a remote to the person wearing white shoes, the agent separates the seated target from a standing bystander in gray from the first frame and stops in front of the correct one. The one weak point is that the small remote falls below the low camera, so the pickup is assumed rather than confirmed. EMBODIMENT enter a room, recover at the door · real ep… view at source ↗
Figure 10
Figure 10. Figure 10: Entering a room with a doorway recovery, success. real ep4. The agent finds the kitchen on the right and, at the narrow doorway, reads a body catch correctly as its own unseen rear, frees itself with small pivots, and stops inside. This is the positive counterpart to the exit failure in fig. 12. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Exit through a doorway, control, success. go2 ep7. Same start and instruction as fig. 12. The agent drives fully out of the door before turning, so the turn is clean and it walks straight to the cutout. EMBODIMENT exit-door embodiment · go2 ep6 “Exit through the door, turn left, and walk to the cardboard cutout of a person.” Turns left the moment the camera clears the doorway — but the ~50 cm body behind … view at source ↗
Figure 12
Figure 12. Figure 12: Exit through a doorway, body catch, failure. go2 ep6. The agent turns while still in the opening, before the body behind the lens has cleared it. The rear jams the wall edge, it misreads its pose as facing back into the room, and it never reaches the target. It does not leave room for its own body to turn. jams the wall edge, the agent then misreads its own pose as facing back into the room it just left, … view at source ↗
Figure 13
Figure 13. Figure 13: Counting gate, failure. real ep3. Asked to pass two pillars and turn, the agent reads the corridor correctly but never fixes the count. It relabels the second pillar across frames and turns only at the end wall, one pillar too far. MULTI-GOAL multi-goal (long-horizon) · go2 ep17 “Walk to the yellow tape on the floor, then turn around and walk to the robot arm on the table.” Leg 1 (yellow tape) is clean. A… view at source ↗
Figure 14
Figure 14. Figure 14: Long route with a reference flip, failure. go2 ep17. The first leg to the tape is clean. After the agent turns around, its memory of the target seen earlier decays, and it stops at a monitor cart that it reports as the robot arm. The clearest terminal error of the set. on a heuristic that a floor target which drops below the low camera must lie underfoot, which stops it early or late whenever the geometry… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 16 linked inside Pith

  1. [1]

    Zhang, K

    J. Zhang, K. Wang, R. Xu, G. Zhou, Y . Hong, X. Fang, Q. Wu, Z. Zhang, and H. Wang. Navid: Video-based vlm plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    M. Wei, C. Wan, X. Yu, T. Wang, Y . Yang, X. Mao, C. Zhu, W. Cai, H. Wang, Y . Chen, et al. Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    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, volume 38, pages 7641–7649, 2024

    2024

  4. [4]

    J. Chen, B. Lin, R. Xu, Z. Chai, X. Liang, and K.-Y . Wong. Mapgpt: Map-guided prompting with adaptive path planning for vision-and-language navigation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9796–9810, 2024

    2024

  5. [5]

    R. Gong, Y . Guo, J. Hu, J. Kong, X. Leng, T. Li, W. Li, F. Liu, Z. Liu, J. Lu, M. Luo, C. Ming, Y . Shen, J. Tao, Z. Wang, M. Yin, et al. ABot-N1: Toward a general visual language navigation foundation model, 2026. URL https://arxiv.org/abs/2607.10383

    Pith/arXiv arXiv 2026

  6. [6]

    M. Wei, C. Wan, J. Peng, X. Yu, Y . Yang, D. Feng, W. Cai, C. Zhu, T. Wang, J. Pang, and X. Liu. Ground slow, move fast: A dual-system foundation model for generalizable vision-and-language navigation, 2025. URL https://arxiv.org/abs/2512.08186

    arXiv 2025

  7. [7]

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao. ReAct: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  8. [8]

    C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. R. Narasimhan. SWE-bench: Can language models resolve real-world github issues? InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=VTF8yNQM66

    2024

  9. [9]

    J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. Narasimhan, and O. Press. SWE-agent: Agent-computer interfaces enable automated software engineering. InNeurIPS, 2024

    2024

  10. [10]

    Anderson, Q

    P. Anderson, Q. Wu, D. Teney, J. Bruce, M. Johnson, N. Sünderhauf, I. Reid, S. Gould, and A. Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3674–3683, 2018. 12

    2018

  11. [11]

    Krantz, E

    J. Krantz, E. Wijmans, A. Majumdar, D. Batra, and S. Lee. Beyond the nav-graph: Vision-and- language navigation in continuous environments. InEuropean Conference on Computer Vision, pages 104–120. Springer, 2020

    2020

  12. [12]

    Y . Qiao, W. Lyu, H. Wang, Z. Wang, Z. Li, Y . Zhang, M. Tan, and Q. Wu. Open-nav: Exploring zero-shot vision-and-language navigation in continuous environment with open-source llms. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 6710–6717. IEEE, 2025

    2025

  13. [13]

    X. Shi, Z. Li, W. Lyu, J. Xia, F. Dayoub, Y . Qiao, and Q. Wu. Smartway: Enhanced waypoint prediction and backtracking for zero-shot vision-and-language navigation. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 16923–16930. IEEE, 2025

    2025

  14. [14]

    Y . Li, C. Li, H. Shi, J. Luo, J. Cai, M. Yang, and T. Qin. AgenticNav: Zero-shot vision-and- language navigation as a tool-calling harness, 2026. URL https://arxiv.org/abs/2606.10577

    Pith/arXiv arXiv 2026

  15. [15]

    Zhang, G

    J. Zhang, G. Zhou, H. Yin, Y . Huang, Z. Lei, Q. Peng, H. Yuan, J. Zhang, X. Guo, X. Chen, A. Yang, F. Huang, J. Lin, D. Liu, J. Zhou, Z. Yu, J. Fan, Z. Liang, P. Lin, Y . Wang, A. Chen, K. Yan, X. Xu, J. Li, L. Hu, M. Zhang, S. Li, W. Xiao, S. Bai, X. Ren, C. Lv, C. Wu, and X.-H. Chen. Qwen-RobotNav technical report: A scalable navigation model designed ...

    Pith/arXiv arXiv 2026

  16. [16]

    Zhang, A

    J. Zhang, A. Li, Y . Qi, M. Li, J. Liu, S. Wang, H. Liu, G. Zhou, Y . Wu, X. Li, Y . Fan, W. Li, Z. Chen, F. Gao, Q. Wu, Z. Zhang, and H. Wang. Embodied navigation foundation model, 2025. URL https://arxiv.org/abs/2509.12129

    arXiv 2025

  17. [17]

    Driess, F

    D. Driess, F. Xia, M. S. M. Sajjadi, C. Lynch, A. Chowdhery, B. Ichter, A. Wahid, J. Tompson, Q. Vuong, T. Yu, W. Huang, Y . Chebotar, P. Sermanet, D. Duckworth, S. Levine, V . Vanhoucke, K. Hausman, M. Toussaint, K. Greff, A. Zeng, I. Mordatch, and P. Florence. PaLM-E: An embodied multimodal language model. InProceedings of the 40th International Confere...

    2023

  18. [18]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florence, C. Fu, M. G. Arenas, K. Gopalakrishnan, K. Han, K. Hausman, A. Herzog, J. Hsu, B. Ichter, A. Irpan, N. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, I. Leal, L. Lee, T.-W. E. Lee, S. Levine, Y . Lu, H. Michalewski, I. Mordatch, K. Pe...

    2023

  19. [19]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. OpenVLA: An open-source vision-language-action model, 2024

    2024

  20. [20]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky.π0: A vision-language- action flow model for general robot control, 2024

    2024

  21. [21]

    Y . Long, X. Li, W. Cai, and H. Dong. Discuss before moving: Visual language navigation via multi-expert discussions. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 17380–17387. IEEE, 2024. 13

    2024

  22. [22]

    Y . Long, W. Cai, H. Wang, G. Zhan, and H. Dong. Instructnav: Zero-shot system for generic instruction navigation in unexplored environment. InConference on Robot Learning, pages 2049–2060. PMLR, 2025

    2049

  23. [23]

    D. Li, W. Chen, and X. Lin. TINA: Think, interaction, and action framework for zero-shot vision language navigation, 2024. URL https://arxiv.org/abs/2403.08833

    Pith/arXiv arXiv 2024

  24. [24]

    Jeong, G.-C

    S. Jeong, G.-C. Kang, J. Kim, and B.-T. Zhang. Zero-shot vision-and-language navigation with collision mitigation in continuous environment, 2024. URL https://arxiv.org/abs/2410.17267

    Pith/arXiv arXiv 2024

  25. [25]

    H. Li, X. Dong, H. Jiang, Y . Zhou, and X. Ma. CMMR-VLN: Vision-and-language navigation via continual multimodal memory retrieval, 2026. URL https://arxiv.org/abs/2603.07997

    arXiv 2026

  26. [26]

    Z. Li, S. Li, Z. Zhang, B. Li, and S. Zhou. DV-VLN: Dual verification for reliable LLM-based vision-and-language navigation, 2026. URL https://arxiv.org/abs/2601.18492

    arXiv 2026

  27. [27]

    D. An, H. Wang, W. Wang, Z. Wang, Y . Huang, K. He, and L. Wang. Etpnav: Evolving topolog- ical planning for vision-language navigation in continuous environments.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  28. [28]

    Bjorck, Z

    J. Bjorck, Z. Li, Y . Man, J. Wang, A.-C. Cheng, S. Liu, S. Wang, Z. Yu, A. Badki, S. Birchfield, V . Blukis, Y . Chebotar, S. Chen, S. Leng, Y .-C. Chou, T. Ding, B. Li, Z. Luo, H. Su, J. Tremblay, T. Wang, B. Wen, J. Wu, X. Xie, H. Ye, H. Yin, K. R. Zentner, L. Gui, Y .-X. Wang, Y . Zhu, L. Fan, and J. Kautz. Vesta: A generalist embodied reasoning model...

    Pith/arXiv arXiv 2026

  29. [29]

    J. Zhou, S. Lin, J. Li, S. Fu, G. Zhou, and Q. Wu. Automating the design of embodied agent architectures, 2026. URL https://arxiv.org/abs/2606.30111

    Pith/arXiv arXiv 2026

  30. [30]

    mini-swe-agent: The minimal AI software engineering agent

    The SWE-agent team. mini-swe-agent: The minimal AI software engineering agent. https: //github.com/SWE-agent/mini-swe-agent, 2025

    2025

  31. [31]

    G. Wang, Y . Xie, Y . Jiang, A. Mandlekar, C. Xiao, Y . Zhu, L. Fan, and A. Anandkumar. V oyager: An open-ended embodied agent with large language models.Transactions on Machine Learning Research, 2024

    2024

  32. [32]

    Savva, A

    M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Malik, et al. Habitat: A platform for embodied ai research. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9339–9347, 2019

    2019

  33. [33]

    Cheng, Y

    A.-C. Cheng, Y . Ji, Z. Yang, Z. Gongye, X. Zou, J. Kautz, E. Bıyık, H. Yin, S. Liu, and X. Wang. NaVILA: Legged robot vision-language-action model for navigation, 2024. URL https://arxiv.org/abs/2412.04453

    Pith/arXiv arXiv 2024

  34. [34]

    Z. Wang, X. Yu, Y . Rao, Y . Ling, Y . Li, O. Wang, M. Gao, Y . Zhou, Y . Liang, Z. Liu, Y . Zhang, R. Huang, X. Xu, B. Yuan, Y . Yuan, X. Tan, H. Zhang, Y . Huang, S. Zhang, H. Wu, H. Hu, and Z. Zhang. Hy-Embodied-VLM-1.0: Efficient physical-world agents, 2026. URL https: //arxiv.org/abs/2607.12894

    Pith/arXiv arXiv 2026

  35. [35]

    X. Xue, J. Hu, M. Luo, S. Xie, J. Chen, Z. Xie, K. Quan, W. Guo, M. Xu, and Z. Chu. OmniNav: A unified framework for prospective exploration and visual-language navigation, 2025. URL https://arxiv.org/abs/2509.25687

    arXiv 2025

  36. [36]

    Claude agent SDK

    Anthropic. Claude agent SDK. https://docs.anthropic.com/en/api/agent-sdk, 2025

    2025

  37. [37]

    Unrolling the Codex agent loop

    OpenAI. Unrolling the Codex agent loop. https://openai.com/index/ unrolling-the-codex-agent-loop/, 2026. 14

    2026

  38. [38]

    S. Lin, Z. Li, X. Zhao, G. Zhou, L. Wang, R. Wei, R. Tang, J. Li, H. Wang, J. Pang, A. van den Hengel, J. Liu, and Q. Wu. VLNVerse: A benchmark for vision-language navigation with versatile, embodied, realistic simulation and evaluation, 2025. URL https://arxiv.org/abs/2512. 19021

    2025

  39. [39]

    Y . Hong, Z. Wang, Q. Wu, and S. Gould. Bridging the gap between learning in discrete and continuous environments for vision-and-language navigation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15439–15449, 2022

    2022

  40. [40]

    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), pages 4392–4412, 2020

    2020

  41. [41]

    Y . Xu, D. Lu, Z. Shen, J. Wang, Z. Wang, Y . Mao, C. Xiong, and T. Yu. Agenttrek: Agent trajectory synthesis via guiding replay with web tutorials.arXiv preprint arXiv:2412.09605, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Y . Qin, Y . Ye, J. Fang, H. Wang, S. Liang, S. Tian, J. Zhang, J. Li, Y . Li, S. Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents.arXiv preprint arXiv:2501.12326, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Y . Wei, O. Duchenne, J. Copet, Q. Carbonneaux, L. Zhang, D. Fried, G. Synnaeve, R. Singh, and S. I. Wang. Swe-rl: Advancing llm reasoning via reinforcement learning on open software evolution.arXiv preprint arXiv:2502.18449, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022

    2022

  45. [45]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, and D. Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  46. [46]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Chang, A

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang. Matterport3D: Learning from RGB-D data in indoor environments. InInternational Conference on 3D Vision, pages 667–676, 2017. doi:10.1109/3DV .2017.00081

    arXiv 2017

  48. [48]

    Ollama documentation

    Ollama. Ollama documentation. https://docs.ollama.com/, 2026

    2026

  49. [49]

    {instruction}

    Alibaba Cloud. DashScope documentation. https://help.aliyun.com/zh/model-studio/, 2026. A Experimental Settings This appendix records the complete experimental configuration. The authoritative source is the released run registry (configuration as code, section 4.1); this appendix is a faithful transcription of it. A.1 Task, benchmark, and metrics Benchmar...

    2026