VLN Support Status

This page does two jobs. Part I — Support answers which VLN methods AgentCanvas runs or wants to run: a single support matrix (shipped · in-progress · planned · wanted), then the env / tool / method machinery underneath. Part II — Landscape places those methods in the wider VLN literature — the representative and most-cited methods, and the 2025–2026 frontier — so the supported subset is legible against the whole field. Full per-era method narrative lives in the research survey VLN — Methods; this page is the status + map view.

Part I — Methods we support

1. Support matrix

Every VLN method AgentCanvas ships, is building, or wants to build. This is the single source of truth for "what we support" — the env / tool / method backing is in §2–§4, the wider field is Part II. ⭐ marks the PortBench v1 / AAS-seed set.

Today's load-bearing demo set is the ✅ rows — mapgpt_mp3d, smartway_ce, smartway_mono_ce (+ navgpt_mp3d on gpt-4). The ✓ in-progress rows run end-to-end but aren't paper-comparable yet; the ⏳ unverified rows are not a code-quality verdict — most are prompt / tool-wiring / config issues or model-tier gaps (gpt-5-mini vs gpt-5-full) not yet chased down. Verified graphs live under workspace/graphs/vln/verified/, the rest under vln/unverified/ (bare-stem resolution still works via WorkspaceComponentRegistry.resolve_graph_path).

2. Environment nodesets

VLN simulators wrapped as server-mode BaseNodeSets (own subprocess + Python env).

2.1 Habitat (VLN-CE)

• File: workspace/nodesets/env/env_habitat.py; pinned to the ac-vlnce conda env (Python 3.8) by the habitat-sim 0.1.7 wheel.
• Datasets: R2R-CE today; RxR-CE planned (E2).
• Env panel: HabitatEnvPanel — set_split / set_episode / peek_episode (ADR-server-002).
• Action space: Discrete(4) — STOP / FORWARD / LEFT / RIGHT.
• VRAM: actual peak in 20-worker replicated runs ~22.6 GB / 24 GB; the scheduler's 8000 MB declaration is ~3× conservative.

2.2 Matterport3D (MP3D)

• File: workspace/nodesets/env/env_mp3d/__init__.py — MatterSim discrete panoramic navigation.
• Datasets: R2R · R4R · RxR · REVERIE · CVDN / NDH — all six schemas behind one nodeset via the dataset → split → episode_index cascade in MP3DEnvPanel.
• Action space: dynamic-cardinality discrete viewpoints (viewpoint_id TEXT + cand_vpids_json) — unified with Habitat's once TODO #46 lands.

3. Perception & tool nodesets

Method-side nodesets that expose tools any VLN agent can wire — they don't pick a method themselves.

4. Method nodesets

Reasoning-side nodesets implementing a specific VLN paper / family. (Status per §1.)

4.1 NavGPT — nodesets/method/navgpt.py

LLM-as-reasoning-core: action parsing, history management, output formatting. Pairs with navgpt_mp3d.json (discrete) and navgpt_ce.json (CE). NavGPT-MP3D is verified with gpt-4 — only gpt-4 yields a non-zero SR; gpt-4o over-confidently STOPs on long ReAct prompts and collapses to SR=0. NavGPT-CE remains SR=0.

4.2 MapGPT — nodesets/method/mapgpt.py ✅

ACL 2024 — linguistic topological map inside the LLM prompt + adaptive multi-step planning. First VLN method validated end-to-end on AgentCanvas. Pairs with mapgpt_mp3d.json. Use the MapGPT_72 subset (216 eps) for paper-comparable headlines; paper SR 0.477 (gpt-4v) / 0.463 (gpt-4o).

4.3 Open-Nav — nodesets/method/opennav.py

ICRA 2025 — prompts, parsers, ensemble / fusion / tie-break, history; wired to model_ram + vlm_spatialbot (via opennav__select_candidate_views) + opennav_waypoint on Habitat-CE. Pairs with opennav_habitat.json (currently SR=0; the decision chain stalls after step 1 — pre-existing, the graph was never verified).

4.4 SpatialNav — nodesets/method/spatialnav.py

MP3D discrete — 8-compass observation + SSG enrichment fork of NavGPT-MP3D. Platform-internal extension, not a paper port. ("Pre-exploration via SLAM" is in fact MP3D ground-truth .house + _semantic.ply, not SLAM.) Pairs with spatialnav_mp3d.json.

4.5 Policy CMA — nodesets/policy/policy_adapter_vlnce/ (+ env/env_adapter/)

Cross-Modal Attention VLN-CE baseline, run through the General Policy Adapter pipeline (env_adapter stages 1/5 + policy_adapter_vlnce stages 2/3/4; variant CMA_PM_DA_Aug). Hidden state on explicit hidden_in / hidden_out ports. Pairs with vln/verified/straightforward.json. The original single-node policy_cma.py port was deleted 2026-07-04 (superseded; git history keeps it).

4.6 SmartWay / SmartWay-mono — nodesets/method/smartway/ + smartway_mono/ ✅

VLN-CE waypoint-predictor family. smartway dual-frame, smartway_mono single-frame. smartway-mono SR 0.270 vs paper 0.29 after two 2026-05-16 fixes (place_id + image_labels wiring); smartway-CE author-verified 2026-06-01. Default planner LLM gpt-5-mini (needs temperature=1.0 + max_tokens=2000). Pairs with smartway_ce.json, smartway_mono_ce.json.

4.7 DiscussNav — nodesets/method/discussnav.py ✓

MP3D multi-agent discussion — 8-expert LLM fan-out → aggregator. Pairs with discussnav_mp3d.json; fitness not yet driven to a paper-comparable number. Exercises F3 parallel execution.

4.8 AO-Planner — nodesets/method/aoplanner/ ✓

AAAI 2025 — Affordances-Oriented Planning: two-VLM Visual Affordances Prompting (VAP) over SAM ground masks + projected waypoints, with a SmartWay-cloned action decider. Faithful pure-foundation-model port — the released code's supervised ETPNav TRM waypoint predictor + ZeroShotGraphMap alignment are intentionally dropped. Uses model_grounding_dino for affordance grounding. Pairs with aoplanner_ce.json; runs end-to-end — headline SR 0.10 / oracle 0.20 @20ep (noisy: gpt-5-mini is forced to temp=1.0, so run-to-run SR variance at N≤20 exceeds any tuning delta; the multi-hop controller has hit SPL 0.899 on individual episodes; paper-comparable headline needs gpt-5-full + the 216-ep MapGPT72 slice) (M7). See the AO-Planner nodeset page.

4.9 Three-Step Nav — nodesets/method/threestepnav/ ✓

Zero-shot LLM-VLN (arXiv 2604.26946) — a descendant of Open-Nav (its agent class literally is class Open_Nav), so it reuses the exact env_habitat + opennav_waypoint + RAM/SpatialBot substrate (generic model_ram / vlm_spatialbot wrappers since 2026-07-04) and adds the three method-side steps: global sub-instruction decomposition → local waypoint choice → back-check verify. Ported to byte-level fidelity over five cold audit passes (82-check upstream-import harness; waypoint path tensor-equal 6/6; perception layer aligned 2026-07-03 incl. the 1024 px render and a previously never-loaded DDPPO depth encoder). The aligned-stack gpt-5 run scores SR* 0.24 vs the paper's 0.34 at 25 ep — statistically compatible — with TL 9.25 ≈ paper 9.18; the official stop-gated SR is 0.08 because upstream's reported SR has no STOP requirement (two-ruler analysis in the nodeset page §4.1). A 100-episode gpt-5-mini scale check lands SR* 0.29 · OSR 0.39 (equal to the paper). Single-version monolith threestepnav (the Phase-2 decomposition was removed 2026-07-03); graph vln/verified/threestepnav_ce.json.

4.10 VLN-CE policy registry — nodesets/policy/policy_adapter_vlnce/

A 12-variant R2R-CE baseline registry (CMA × 7 + Seq2Seq × 5, from the VLN-CE paper's Table 1) selectable via a variant ConfigField. Upstream published only 2 checkpoints — CMA_PM_DA_Aug (SPL 0.30) and Seq2Seq_DA (SPL 0.23); the other 10 stay in the registry for topology completeness but their dropdown label carries a "(not released by paper authors)" suffix so a run fails fast with a clear missing-file reason. This registry is what §4.5 selects from.

5. Datasets, action spaces, metrics

Datasets: R2R · R4R · RxR · REVERIE · CVDN/NDH (MP3D discrete); R2R-CE shipped, RxR-CE planned (E2); VLN-CE is the continuous family. Split note: R2R val_unseen front-loads easy scans — use MapGPT_72 (216 ep) for paper-comparable headlines; NavGPT batch default 10 ep, stride=3.

Action spaces (env-specific until TODO #46): Habitat Discrete(4) (0=STOP/1=FWD/2=LEFT/3=RIGHT); MP3D dynamic-cardinality viewpoints (viewpoint_id + cand_vpids_json). Metrics: SPL (primary), SR (within 3 m), nDTW, SDTW.

Part II — Field landscape

6. Representative & most-cited

Where the supported methods (§1) sit in the wider VLN literature. Citation counts are a Semantic Scholar snapshot, 2026-06-14; Time is arXiv first-submission YY.MM; the Ours badge ties each row back to §1 status. A code link marks a method with a verified public GitHub repository (link-check snapshot 2026-06-20); no badge = no public code found, and code soon = official repo exists but code not yet released. Rows are ordered by arXiv date (oldest first). Full per-era narrative: VLN — Methods.

Single-box trained transformers (pretraining-era, HAMT/DUET) are deliberately low-priority to port — under PortBench rules they collapse to one LLMCall/policy node. The methods we invest in are the graph-shaped LLM agents (⭐).

7. Frontier (2025–2026, zero-shot LLM)

Newest zero-shot LLM-VLN — high direction-signal, citations still accruing, none ported (📚). Sorted by arXiv date. A code link marks the few with a verified public repo (snapshot 2026-06-20); most have no public code released yet. Detail in survey §2.5.