Method NodeSets
Agent and method implementations. Reasoning-only nodesets consume env and foundation-model tools over wires; neural-policy baselines run through the General Policy Adapter pipeline — one general env_adapter nodeset plus a per-domain policy adapter (policy_adapter_vla, policy_adapter_vlnce), composed with the env nodeset at run time.
| NodeSet | File | Description |
|---|---|---|
| MapGPT | mapgpt.py | MapGPT (ACL 2024) linguistic topo-map + adaptive planning — 8 nodes on env_mp3d waypoint nav |
| DiscussNav | discussnav.py | DiscussNav (ICRA 2024) multi-expert-discussion port — 9 nodes + RAM/InstructBLIP on env_mp3d. Faithful (justified deviations) / verified (smoke): SR 0.20, 5-ep gpt-5-mini |
| SmartWay | smartway/ + 2 server | SmartWay (IROS 2025) decomposed port — TRM waypoints + RAM+ tags + backtracking return on env_habitat hightolow |
| AO-Planner | method/aoplanner/ + 1 server | AO-Planner (AAAI 2025) two-VLM Visual Affordances Prompting, pure-FM framing — 15 nodes, 4 perception lanes + multi-point Path execution (env_habitat step_path). Faithful (justified deviations) / verified at 100-ep scale: run-config chain matched 2026-07-04 (512×512 render · horizon 5) → SR 0.35 · OSR 0.50 · SPL 0.21 (rand100, gpt-5-mini both VLMs, run 20260704_011258); not paper-comparable (model substitution); graph in vln/verified/ |
| Three-Step Nav | method/threestepnav/ | Three-Step Nav (arXiv 2604.26946) global decompose + image navigator + back-check judge on env_habitat hightolow — reuses Open-Nav's TRM + RAM + SpatialBot-3B. Faithful (justified deviations) — reproduced under the paper's own metric: five cold audit passes closed every port-side gap (82-check upstream-import byte harness; waypoint path 6/6 tensor-equal); the aligned-stack gpt-5 run (1024 px) scores SR* 0.24 vs paper 0.34 — compatible at n=25 — with OSR 0.28 · TL 9.25 ≈ paper; official stop-gated SR 0.08 (upstream's reported SR has no STOP gate — definitions on the page, §4.1); 100-ep mini-tier scale check SR* 0.29 · OSR 0.39 (= paper). Single-version monolith (Phase-2 decomp removed 2026-07-03); graph in vln/verified/ |
| NavGPT | navgpt_mp3d_tools.py + navgpt.py | NavGPT (AAAI 2024) GPT-4 ReAct loop on env_mp3d (paper caches) + the in-house NavGPT-CE variant |
| ExploreEQA | method/explore_eqa.py | Explore until Confident (Ren et al. 2024) — TSDF frontier exploration + Prismatic token-likelihood scoring + post-hoc vote on env_hmeqa |
| VoxPoser | method/voxposer/ | VoxPoser (Huang et al. 2023) decomposed port — 7 LMPs + value-map planning on a GT snapshot, closed-loop waypoint follow on env_libero |
| General Policy Adapter | env/env_adapter/ + policy/policy_adapter_vla/ + policy_adapter_vlnce/ | The adapter paradigm and both instances — env-adapter ↔ CanonicalDict ↔ model-adapter, four stages + one model-forward, three-nodeset composition, filename discovery. Covers the VLA stack (Pi0/SmolVLA/DP/RT-1-X) and the VLN-CE baselines (CMA/Seq2Seq, verified on straightforward.json) |