Embodied Navigation Weekly (2026-10-02 to 2026-10-08)

2026-10-11
  • VLN
  • VLA
  • Embodied Navigation
  • Weekly Digest
  • arXiv

1. Key conclusions

  • Of 147 newly collected works, 3 explicitly report R2R-CE results, but their numbers measure different things. NavGPT-3 reports SR of 81.51 / 90.43 on continuous R2R-CE / RxR-CE; StageVLN reports 56.3% SR and 51.4% SPL on R2R-CE val-unseen. PG-VP’s 84.9% / 83.2% measure guidance toward intended low-risk actions on these continuous benchmarks, with navigation SR falling by 6.8 / 7.9 percentage points. Our assessment: align tasks, splits, and safety objectives before comparing performance. These figures do not form a common leaderboard, and none refers to discrete R2R.
  • A central runtime question is becoming who takes control of motion, and when. NavGPT-3 schedules separate reasoning, acting, and monitoring threads; SuperNav connects a generalist model to navigation tools through visual points. In simulations that keep evolving during inference, RT-SAFE finds 12.3 times as many collisions under real-time execution as in matched static evaluation. Our assessment: correct answers still require timely execution, interruption, and recovery; task completion alone does not establish closed-loop reliability.
  • Long-term memory now addresses both where evidence was observed and how it changes the next decision. LiSoNav / IVAM-Nav anchors small-object memory to observation viewpoints and tests revalidation after relocation; MarvisNav displays exploration states directly on visual route candidates. Our assessment: for ground robots, reducing repeated search and correcting stale beliefs are more useful tests than the number of frames a context can hold.
  • Compute reduction must preserve decision-relevant information and be measured in the complete loop. StageVLN confines geometric auxiliaries to training; LiteNWM scores candidate trajectories using future latent representations; LaTraNav couples slow semantics with fast planning. Our assessment: training-time structural supervision, inference-time predictive compression, and asynchronous scheduling optimize different parts of a system. RTX 5090 speedups do not directly establish onboard performance.

This issue adds 147 records corresponding to 147 independent arXiv works, published between 2026-10-02 and 2026-10-08. Another 5 existing papers received version updates and are excluded from the new-work count. There are 0 new WeChat articles. Automatic primary-pool assignments comprise 32 navigation and 17 embodied-agent works, totaling 49; the secondary pool comprises 87 VLA/manipulation, 1 autonomous-driving, and 10 other works, totaling 98. Automatic labels contain boundary errors: navigation works such as PG-VP and SpikingVLA fall in the secondary pool, while non-embodied work such as IndexAct enters the primary pool. Reading priorities below follow the actual research content; these counts retain the automatic classification. The categorized index covers all 147 works.

2. Priority reading list

  1. NavGPT-3 · Continuous R2R-CE / RxR-CE and runtime scheduling
    • Contribution: separate reasoning, action, and monitoring contexts with interruption and motion-control handoffs.
    • Evidence: full-system SR of 81.51 / 90.43 in the abstract; corresponding splits are unspecified and require protocol checks before comparison.
    • Why read: combines continuous VLN with an AgentOS-like runtime.
  2. StageVLN · Training-time spatial supervision for continuous navigation
    • Contribution: geometry, relative heading, and route-progress supervision are training-only.
    • Evidence: a 4B backbone; R2R-CE val-unseen SR 56.3%, SPL 51.4%; RxR-CE SR 54.3%, with its split unspecified in the abstract.
    • Why read: tests representation improvements without adding deployment modules.
  3. PG-VP · Non-visual hazards in continuous navigation
    • Contribution: turns thermal or radiation risk into moving virtual obstacles that steer frozen OmniNav.
    • Evidence: intended low-risk action guidance of 84.9% / 83.2% on R2R-CE / RxR-CE val-unseen, at an SR cost of 6.8 / 7.9 percentage points.
    • Why read: exposes the tradeoff between avoiding risk and reaching the goal; the guidance rate is not SR.
  4. MarvisNav · Zero-shot ObjectNav
    • Contribution: projects candidate topological nodes and exploration states onto the current view.
    • Evidence: HM3D SR 81.2%, SPL 42.5%; the abstract reports VLM calls at 7.5% of WMNav’s.
    • Why read: a concrete memory interface permits controls that hold information content constant.
  5. LiSoNav / IVAM-Nav · Lifelong search with changing object layouts
    • Contribution: starts with empty memory and uses multi-view inspection and viewpoint-anchored memory to find small objects.
    • Evidence: LiSoNav-Eval has 28 indoor scenes and 45 small-object categories; no verifiable SR / SPL values in the abstract.
    • Why read: assumptions match persistent household use.
  6. SuperNav · General requests in unfamiliar environments
    • Contribution: skills, tools, and progress management support sustained navigation without navigation-specific MLLM fine-tuning.
    • Evidence: instance-level, multi-object, demand-driven, HM3D, and quadruped evaluations are reported, without absolute metrics.
    • Why read: useful for comparing how runtimes connect general reasoning to specialized execution.
  7. LiteNWM · Candidate-trajectory evaluation for visual navigation
    • Contribution: shared visual encoding and joint multi-horizon future latent prediction for scoring.
    • Evidence: macro-averaged offline trajectory error on RECON / SCAND / SACSoN falls 17.56% relative to NoMaD+NWM-XL, with a 128.00-fold RTX 5090 speedup. Real-robot success rises from 43.3% to 83.3% relative to NoMaD.
    • Why read: check candidate counts, hardware, closed-loop trial scale, and scorer-transfer conditions.

Problem. Long reasoning and low-latency motion require different execution rhythms.

Method. The paper gives reasoning, acting, and monitoring threads separate contexts, tools, and permissions, managed by a scheduler.

Evidence. The abstract reports minimum reaction time falling from 3–19 seconds per language-model decision to 0.5–1 second per action-policy step. Interpret the SR results above with their specific evaluation splits.

Value. Our assessment: interruption, context isolation, and execution-feedback interfaces are the transferable components.

Limits. Minimum reaction time is not worst-case end-to-end latency; benchmark success does not establish general human-level physical competence.

Next step. Inspect motion-control handoffs, monitoring triggers, and recovery logs before choosing a replication scope.

2. StageVLN: learn spatial structure during training

Problem. Action supervision alone may fail to retain geometry, orientation, and global progress.

Method. The paper combines hierarchical supervision from a frozen geometry model with relative-heading and expert-route progress objectives, removing auxiliaries at deployment.

Evidence. R2R-CE val-unseen SR / SPL are listed above; the abstract lacks a matched-backbone baseline and training costs.

Value. Our assessment: suitable when deployment architecture is fixed but training can change.

Limits. Removing auxiliary modules neither reduces the original backbone’s computation nor makes training free.

Next step. Ablate geometry, heading, and progress supervision separately under fixed data and training budgets.

3. MarvisNav: make memory part of visual route selection

Problem. Separate textual history forces the model to infer how memory corresponds to available routes.

Method. The paper maintains local exploration progress in a topological graph and displays candidate nodes with their states in egocentric views.

Evidence. The abstract reports HM3D SR / SPL and fewer calls, plus real-robot validation without a trial count.

Value. Our assessment: presentation may drive gains, making an interface-level replication useful.

Limits. Fewer model calls need not reduce total latency; mapping and projection still cost time.

Next step. Hold nodes, exploration information, and model constant while comparing text, a separate map, and image overlays.

4. LiSoNav / IVAM-Nav: interpret non-detection through viewing conditions

Problem. Small objects are occluded and moved; last-known positions cannot support sustained search alone.

Method. The paper inspects supporting surfaces from complementary views and associates memory with observation viewpoints for reuse and revalidation.

Evidence. The benchmark separates unchanged and relocated targets and requires empty initial scene memory. Scale is listed above; the abstract gives no performance values.

Value. Our assessment: search efficiency, visibility, and stale memory are tested together.

Limits. The entry cannot establish perception cost, sensitivity to pose error, or real-robot reproducibility.

Next step. Record old-location rechecking cost, useful observations, and recovery time after relocation separately.

5. SuperNav: inspect how general requests become executable targets

Problem. Navigation-specific fine-tuning may constrain generalization to new requests and environments.

Method. The paper retains a generalist MLLM and adds skills, tools, context management, and a shared visual-point interface.

Evidence. The abstract reports gains over four baselines on multiple tasks and quadruped deployment, without comparable absolute values.

Value. Our assessment: the semantic-to-executable-target interface deserves close inspection.

Limits. “Any task in any scene” is the title, not evidence of unlimited generalization.

Next step. Test feedback and replanning for unknown objects, ambiguous requests, and unreachable targets, beyond successful demonstrations.

4. Transferable methods

  • Intermittent perception: ALONE. Propagates spatial beliefs through actions and looks again when reliability is inadequate and observation would help. Success is 98% / 97% in two drone simulation families; median new-depth observation fractions are 0.9% / 1.3% among successful trials only. Our assessment: useful for ground robots sharing a camera, but wheeled dynamics, moving obstacles, and observation demand in failed trials require new tests.
  • Compact history: LightVLN. One token per historical frame and local aggregation reduce inputs. Reported Orin NX 16 GB rates are 14.61 Hz inference and 11.13 Hz end-to-end updates in real-to-sim hardware-in-the-loop evaluation. Our assessment: borrow the compression mechanism; this is not evidence of autonomous real-world flight or ground-VLN success.
  • Causal-memory evaluation: EMBER-Bench. Tests both next-action selection and causal traceback. The best of 16 models reaches 61.2% accuracy versus 98.3% averaged over two human evaluators. Our assessment: introduce historical constraints such as a previously closed door or relocated goal in ground navigation, and test whether remembered causes change actions correctly. Question answering is still distinct from closed-loop control.
  • Subtask context: RobotUse. The backend handles geometry, planning, and control; subtasks retain detail and return decision-relevant information, while execution updates a persistent playbook. RoboLab task success is 45%, 6.7 percentage points above CaP-X. Our assessment: a useful pattern for navigation subgoal handoffs, subject to new tests of motion-time latency and continuous-state consistency.

5. Categorized index

A = directly relevant to ground navigation; B = a clear transferable mechanism or evaluation contribution; C = low-relevance field observation. Manipulation papers are grouped by theme. Editorial placement does not change automatic pool counts.

5.1 Ground VLN / ObjectNav / semantic navigation

  • LiteNWM (A): Scores trajectories through future latents; see priority list.

  • LaTraNav (A): Asynchronous slow VLM and fast planner; 6.05-fold faster path updates at the same semantic-update rate in the abstract.

  • SuperNav (A): General requests connected to navigation tools; see featured analysis.

  • TAPNAV (A): Actively touches structures for localization and route planning when vision is unavailable.

  • AirGroundVLN (A): Air–ground goal navigation with cross-view memory and regional-to-local planning.

  • LiSoNav / IVAM-Nav (A): Lifelong small-object search under relocation; see featured analysis.

  • SpikingVLA (A): Asynchronous spiking inference; navigation SR / SPL are mentioned without a named benchmark, preventing direct comparison.

  • MixVPR Teach-and-Repeat (A): MixVPR place recognition reduces hardware demands for teach-and-repeat navigation.

  • SiGNgapore (A): Sign-based navigation data from real public spaces for sequential sign-following decisions.

  • NavGPT-3 (A): Navigation thread scheduling and interruption; see featured analysis.

  • COOL (A): Infers ownership from human–object interactions and actively refreshes memory for personalized object finding.

  • RT-SAFE (A): Simulation continues during inference, separating completion from safety events.

  • CUSP (A): Uses annotated risk onset and accumulated alarms for off-road hazards; intervention time is distinct from hazard onset.

  • PG-VP (A): Dynamic prompts for non-visual hazards; see priority list. Automatically labeled VLA/manipulation.

  • Risk-Sensitive Crowd Navigation / AECP (A): Directional uncertainty ellipsoids and tail-risk constraints address shifts in pedestrian motion.

  • Semantic-Aware Humanoid Navigation (A): Addresses hazards missed by elevation maps and command–execution mismatch; distinguish simulated navigation from real locomotion validation.

  • Ackermann VLN Sim-to-Real (A): Continuous VLN transfer to an Ackermann platform; no SPL / nDTW values in the abstract.

  • MarvisNav (A): Visual exploration memory; see featured analysis.

  • Dual-VAE Sim-to-Real (A): Dual VAEs align simulation and real features; nearly 91% in the abstract is an image-classification result, not closed-loop navigation SR.

  • JESSI (A): Joint pedestrian perception and social navigation from LiDAR to executable control.

  • StageVLN (A): Training-time spatial and progress supervision; see featured analysis.

  • Tour-guide Social Navigation (A): Social-force model for following a tour-guide robot; the abstract primarily describes an experimental design.

  • TACET (A): Joint control of personal space and locomotion noise extends social navigation beyond distance keeping.

5.2 Memory, maps, planning, and evaluation

  • 2DGS-Planner (B): Rasterization queries traversable geometry in Gaussian maps; a geometric planning component.

  • Mine Odyssey (B): Long-horizon spatial exploration in reconstructed Minecraft scenes, distinct from real-robot navigation.

  • Arena 5.0 (A): ROS2 social-navigation simulation and scenario generation.

  • SCOPE (B): Control-ready uncertainty for trajectory diffusion, including crowd-navigation evaluation.

  • Ctrl-CWM (B): World-model planning generates crowds that adapt to specified objectives.

  • iAm.md (B): Deployment evidence and persistent object records support skill feasibility assessment.

  • System Switch (B): Studies when to invoke slow reasoning; no variant exits in closed-loop Doom, so offline gains do not establish task success.

  • ActiveLang (B): Selects informative views through semantic uncertainty for active open-vocabulary 3D mapping.

  • VeriFine (B): Co-evolves policy, curriculum, and verifier to address limitations in feedback quality.

  • SPW-Nav (B): Language-driven streaming panorama generation; video quality is distinct from navigation-policy performance.

  • Evidence-Driven Human-Agent-Robot Teaming (B): Evidence collection through authority-bounded services; evidence consists of scenarios and a hardware-in-the-loop prototype.

  • SpaTime (B): Causal geometry tokens and response-time supervision support streaming spatial reasoning.

  • EMHO (B): Revises a frozen model’s harness using execution traces while managing subtask tradeoffs.

  • Attacca (B): Trains continuous search–approach–interaction in Minecraft, retaining the state left by previous tasks.

  • OntoPlan (B): Symbolic scenes and action preconditions support long-horizon robot planning; automatically labeled other.

  • Inspect Robots (B): Modular infrastructure for defining, executing, and terminating physical evaluations.

  • Embodied Guardrail Benchmark (B): Separates defense effectiveness, false positives on benign tasks, and runtime latency.

  • Direction-Conditioned Policies (B): Conditions goal policies on representation-space direction and distance; not specific to language navigation.

  • EMBER-Bench (B): Joint tests of historical causes and next actions; see transferable methods.

  • RobotUse (B): Subtask context and persistent playbooks; see transferable methods.

  • PreAct-Nav (A): Persistent subgoals and action-conditioned prediction support urban navigation; no absolute metrics in the abstract.

  • ROMA (B): Actively obtains visual, acoustic, tactile, and force evidence through object interactions.

5.3 Embodied VLA / mobile manipulation

5.4 Aerial navigation, driving, and other lower-relevance directions

5.5 News and non-paper items

  • No new WeChat articles or separate news items; all entries above are papers.
  • Runtimes are becoming independently testable research objects. NavGPT-3, SuperNav, and EMHO examine scheduling, execution interfaces, and experience-driven harness revision respectively. Our assessment: system organization needs separate ablations under the same underlying model; improvements cannot all be attributed to model capability.
  • Observation, memory, and action need a verifiable causal chain. IVAM-Nav, MarvisNav, and EMBER-Bench address visibility, spatial presentation, and historical constraints. Our assessment: ask which decision memory changed, and whether that change was correct, beyond retrieval and answering scores.
  • Efficiency must be reported alongside safety and completion. The 49 automatically assigned primary works are outnumbered by 98 secondary works; manipulation speedups do not directly establish navigation gains. LiteNWM, LaTraNav, and RT-SAFE motivate measuring complete-loop latency, observation freshness, completion, and collisions together.

Research gaps

  • This issue offers no directly aligned continuous-VLN comparison covering splits, observations, interventions, compute budgets, and dynamic environments. Minimum response time especially must not be treated as a worst-case guarantee.
  • The abstracts do not jointly validate viewpoint-anchored memory and visual route memory under simultaneous pose drift, object relocation, and perception errors.
  • Failures caused by risk avoidance need evaluation beyond SR to distinguish reasonable refusal, necessary detours, and policy degradation.

High priority

  • Read NavGPT-3 and StageVLN closely. Check R2R-CE / RxR-CE splits, training data, inputs, and interruption mechanisms before selecting baselines; do not mix in discrete R2R results.
  • Replicate MarvisNav’s memory-presentation controls. Fix the underlying model and information content; measure revisits, search paths, and actual total latency.
  • Track LiSoNav data and code. Start with empty memory and separately test search efficiency for relocated and unchanged targets.

Medium priority

  • Evaluate PG-VP’s safety–success tradeoff. Keep low-risk action guidance and navigation SR as separate metrics; do not hide costs in a single weighted score.
  • Inspect deployment interfaces in SuperNav, LiteNWM, and LaTraNav. Measure end-to-end latency, observation-to-execution delay, and recovery on the target hardware.
  • Add continuous-world evaluation. Following RT-SAFE, keep pedestrians and obstacles moving during inference and retain matched static controls.

Low priority

  • Defer importing manipulation scores into navigation leaderboards. Transfer aerial navigation, game agents, tabletop manipulation, and driving only with explicit task mappings and validation protocols. Track further materials for work lacking closed-loop metrics.

Feedback