Paper: Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System Authors: Qwen Team — core contributors Jiazhao Zhang*†, Gengze Zhou*†, Hale Yin*, Yiyang Huang*, Zixing Lei*, Qihang Peng*, Haoqi Yuan, Jie Zhang, Xudong Guo, Xiaoyue Chen, An Yang, Fei Huang, Zhibo Yang, Junyang Lin, Dayiheng Liu, Jingren Zhou, Chenxu Lv‡, Chenfei Wu‡, Xiong-Hui Chen†‡ (+16 contributors) Affiliation: Qwen Team, Alibaba Group — part of the Qwen-Robot Suite (with Qwen-RobotManip and Qwen-RobotWorld) arXiv: 2606.18112 · v1 Jun 2026, v3 Jun 29, 2026 (cs.RO) Code: github.com/QwenLM/Qwen-RobotNav · Blog: qwen.ai/blog?id=qwen-robotnav Weights: the shared suite README states there is no plan to release model weights for Qwen-RobotManip or Qwen-RobotNav

Companion reviews: Qwen Team's VLA Program · Qwen-RobotManip · Qwen-RobotWorld · Qwen-VLA (whose navigation results this model supersedes) · NavFoM (the prior nav-foundation-model from the same first author's lineage).

1. The central reframe: multi-task navigation is a context-modeling problem, not an architecture problem. Instruction following needs dozens of steps of global history; tracking needs only the last few frames and treats stale history as noise; object search shifts between the two within one episode. Prior unified models (NavFoM's uniform subsampling, ABot-N0's sliding window) hard-code one assumption. Qwen-RobotNav instead exposes a parameterized observation interface — token budget B, temporal decay γ, per-camera weights w_c, frame-sampling mode — that an external agent can reconfigure at inference time, with training-time randomization over all parameters so any configuration works zero-shot.
2. Radically simple architecture. Qwen3-VL (2B/4B/8B) + a 4-layer MLP head that regresses 8 waypoints (x, y, θ) = 24 dims under MSE. Camera identity and temporal order are conveyed by plain vocabulary tokens ("Time step 0 Front View 〈image〉"); embodiment by a prompt preamble ("Imagine you are a robot / a car…"). Zero architectural modification to the backbone; a new platform needs a new prompt template, not new parameters.
3. 15.6M-sample corpus, five task families — instruction following (5.63M), point-goal (984K), object-goal (2.0M, via a novel skeleton-graph exploration-trajectory generator + VLM open-vocabulary goal annotation), tracking (1.49M), autonomous driving (~3.2M) — plus 15% VL co-training (λ = 1.0) and a 40K text-to-video-synthesized subset. Sim renders are made photorealistic with Qwen-Image-Edit style transfer.
4. SOTA across most navigation benchmarks: VLN-CE R2R 72.1% SR / RxR 76.5% SR (panoramic, 8B — +10.4/+12.1 over NavFoM), HM3Dv2 ObjectNav 75.6% SR (RGB-only, on the harder v2), EVT-Bench 90.0% tracking rate, NAVSIM 91.4 PDMS. As the executor in a two-tier agentic system (Qwen3.6-Plus planner + evidence-notebook memory), it sets EQA SOTA: HM-EQA 76.7 (+7.5), EXPRESS-Bench 79.27 (+10.6), with 77% fewer navigation steps.
5. For the Qwen program: this is the specialist that vindicates the suite decomposition — on the identical VLN-CE Val-Unseen benchmarks, RobotNav-8B beats the generalist Qwen-VLA-Instruct by +14.6 pp on R2R SR (72.1 vs 57.5) and +16.9 pp on RxR SR (76.5 vs 59.6). It also breaks the program's backbone uniformity: RobotNav uses Qwen3-VL, not the flagships' Qwen3.5.

• It gives "agentic navigation" a concrete mechanism. Everyone hand-waves about planners calling navigation policies; this paper specifies the tool signature — W_i = nav_qwennav(L_i, τ_i, Φ_i) with task mode τ ∈ {VLN, PointNav, ObjNav, Tracking} and observation config Φ = (B, γ, {w_c}, m, b_min, b_max) — and shows the planner exploiting it mid-episode (broad-history ObjNav mode for search → recency-focused Tracking mode for approach). The EQA results are the payoff: the same weights, reconfigured per sub-goal, beat systems with dedicated exploration/memory modules while taking 77% fewer steps.
• Training-time randomization as the generalization mechanism. The reason a single model tolerates any inference-time configuration is that it never trains at a fixed one: γ ~ U[1,3], B ~ U[2048,4096], per-camera weights from per-camera ranges, floor/ceiling b_min ~ U_Z[1,8] / b_max ~ U_Z[128,256], frame mode 50/50 random-vs-latest. This is domain randomization applied to the observation encoding rather than the environment — a transferable idea for any long-context embodied model.
• Language-as-structure taken to its logical end. Camera identity, temporal order, and embodiment all enter as ordinary text. The paper even ablates descriptive camera names ("Front Left View") vs numeric azimuths ("right 90 degrees") and finds names win — the pretrained LLM's spatial word semantics do real work. This is the most extreme instance of the Qwen program's language-as-interface doctrine (Review-Qwen-Team-VLA §4).
• First author lineage. Jiazhao Zhang is the first author of NaVid, Uni-NaVid, TrackVLA, and NavFoM — this is the leading VLN-model author building the successor to his own line inside Qwen, and the baselines he beats are largely his own prior systems.

All tasks are unified waypoint trajectory prediction: given instruction L and multi-view stream I over T timesteps × N cameras, predict W = {(x_k, y_k, θ_k)}, K = 8. The challenge: visual tokens scale O(T·N) and different tasks want different subsets of them.

flowchart LR
  subgraph IN["Inputs"]
    OBS["N cameras × T timesteps"]
    PRE["Embodiment preamble<br/>('Imagine you are a robot/car…')"]
    INS["Instruction L"]
    CFG["Config Φ = (B, γ, w_c, m, b_min, b_max)<br/>← set by planner or defaults"]
  end
  CFG --> ALLOC["Task-adaptive token allocation<br/>ω_t = exp(γ·t/(T′−1)) · w_c<br/>constrained floor/ceiling redistribution<br/>token count → per-image pixel resolution"]
  OBS --> ALLOC
  ALLOC --> TAGS["Interleave natural-language tags:<br/>'Time step t' + camera-name + ⟨image⟩"]
  TAGS --> VLM["Qwen3-VL (2B / 4B / 8B)<br/>SigLIP-2 ViT · 2D-RoPE dynamic res<br/>DeepStack multi-level injection"]
  PRE --> VLM
  INS --> VLM
  VLM --> HEAD["4-layer MLP head (512, GELU)<br/>→ 8 waypoints × (x,y,θ) = 24 dims<br/>MSE, per-dataset q99 normalization"]

  classDef v fill:#bbdefb,stroke:#1565c0,color:#000
  classDef a fill:#c8e6c9,stroke:#2e7d32,color:#000
  classDef i fill:#fff9c4,stroke:#f57f17,color:#000
  class VLM v
  class HEAD,ALLOC a
  class IN,TAGS i

• Backbone: Qwen3-VL — SigLIP-2 ViT with native dynamic resolution (2D-RoPE), 2-layer MLP patch merger, DeepStack injection of multi-level ViT features into early LLM layers. The dynamic-resolution ViT is what makes token-budget allocation implementable: an image's allocated token count directly determines its rescaled pixel budget.
• Action head: deliberately minimal (4-layer MLP, hidden 512, GELU) so spatial reasoning stays inside the LLM. Waypoints normalized to [−1,1] by per-dataset 99th-percentile scale factors; plain MSE. No diffusion, no flow matching, no discrete tokens — the lightest action decoder in the entire Qwen robot program.

Four orthogonal control axes:

Allocation: every (timestep, camera) cell gets floor b_min; the remainder distributes proportionally to ω_t·w_c; cells hitting ceiling b_max release surplus iteratively. The allocated token count sets that image's resolution (aspect-ratio-preserving rescale). The authors are candid that this is "an empirically proven heuristic" that a principled allocator could improve.

• Two tiers: an upper planner (Qwen3.6-Plus) does goal decomposition, task-mode selection, and observation-config selection; RobotNav is the reactive executor. Communication is pure natural language.
• Trajectory-to-evidence harness: each rollout returns a compact record (sub-goal, mode, config, progress narrative, salient observations, outcome, key-frame indices) instead of raw streams — key-frame IDs allow later visual recall.
• Two-level memory: per-episode summaries + a persistent evidence notebook holding durable conclusions (searched regions, candidate locations, rejected hypotheses) that survives context compression with auditable belief revisions.
• Auxiliary vision tools (detection, scene understanding, grounding) feed the planner but never predict waypoints.

85% navigation trajectories / 15% VL co-training; loss L = L_traj + λ·L_VL with λ = 1.0 (note: the manipulation flagships use λ = 0.1 — navigation gives language supervision 10× more relative weight).

The corpus source mix is reported as 67% synthetic videos / 32% real-world videos / 1% generated — like Qwen-RobotManip, synthesis-first.

VL co-training (15%): ~1.0M general VL (VQAv2, grounding, multi-image, STEM…) + 873K navigation-specific reasoning — free-form QA at decision points (12 fine-grained action classes) and structured multi-perspective reasoning (History / Scene Analysis / Instruction Progress / Action Reasoning, each becoming an independent QA pair) — + 362K discrete multi-round VLN dialogues (VLN-MME format over CVDN, SOON, REVERIE, SRDF; Matterport3D-rendered panoramas, full decision history kept in-conversation). The stated function goes beyond anti-forgetting: trajectory-only training "collapses into reactive action-sequence mappers", and language-mediated reasoning is framed as a transferable scaffold for VLM-to-VLA transfer.

Optimization: AdamW (β₂ = 0.95, wd 10⁻²), cosine schedule with 3% warmup, LR 2×10⁻⁵ backbone / 1×10⁻⁴ head, grad clip 1.0, batch 256; the 8B model trains in 2,816 H100 GPU-hours — a rare, and notably modest, disclosed compute budget (contrast: neither manipulation flagship discloses any).

Cross-reference to the sibling generalist: Qwen-VLA-Instruct reported 57.5 R2R SR / 59.6 RxR SR on the same Val-Unseen splits — the navigation specialist wins by ~15–17 pp (with the caveat that RobotNav trains on 5.63M instruction-following samples vs Qwen-VLA's 7.5% navigation mixture share).

Also: VLNVerse (full-kinematics physics locomotion) fine-grained 63.75 SR / 57.93 SPL (+12.2 SR / +25.5 SPL over NavFoM — the SPL margins exceeding SR margins indicate markedly more efficient paths); VLN-PE (flash controller) 65.50 SR / 61.19 SPL, best NE (3.73 m) — but note the fall rate: 3.83–4.05% vs 0.22–0.45% for baselines (see §7).

• MP3D / HM3Dv2 (closed-vocab): 4B reaches 52.2% SR on MP3D (best; beats all depth/odometry-using baselines from RGB only) and 75.6% SR on HM3D v2 — evaluated on the harder v2 (216 scenes) while baselines report v1, yet still surpassing Uni-NaVid's v1 73.7%.
• HM3D-OVON (open-vocab): 4B 57.7 / 60.1 / 53.1 SR (Seen/Synonyms/Unseen) — best on two of three splits from a single forward camera, against ABot-N0's panoramic input. SPL trails (24.4 vs 32.1): the skeleton-trained "reach-first" thorough-search behaviour finds more goals via longer paths — a legible data-design artifact.

• EVT-Bench STT: TR 90.0% (4B) / 89.7% (8B) — best among all methods including the dedicated TrackVLA++ (81.0); lowest generalist collision rate (5.70%, 8B). But SR (77.4/78.6) trails specialists ABot-N0 (86.9) and TrackVLA++ (86.0) — the model follows tighter but is more conservative in declaring success.
• Embodied QA (agentic system): Qwen3.6-Plus + RobotNav-8B sets SOTA on all three benchmarks — HM-EQA 76.7 (+7.5 over FAST-EQA), MT-EQA 54.4 (+3.9), EXPRESS-Bench 79.27 LLM score (+10.6) — while cutting normalized steps from 0.65 to 0.15 (the "77% fewer steps" headline).
• NAVSIM navtest: 91.4 PDMS (4B, with 3-frame GT history prior in the prompt) — above NavFoM (+7.1), AutoVLA (+2.3), ReCogDrive, ReflectDrive; NC 99.8, TTC 98.5. Without the history prior: 79.5 PDMS — an 11.9-point swing from prompt-injected ego-history alone.
• AlpaSim (zero-shot closed-loop, at-fault metrics): honest negative result — 0.15/0.17 AlpaSim score vs Alpamayo-R1-10B's 0.72, with 22% close-encounter and 27–34% off-road rates. Open-loop NAVSIM strength does not yet survive zero-shot long-horizon closed-loop transfer.

• Data scaling (12.5% → 100%): strong monotone gains on long-horizon tasks (RxR 52.6 → 69.5; OVON-unseen 37.1 → 53.1; NAVSIM 74.4 → 79.5 w/o hist); tracking saturates early with mild non-monotonicity. Model scaling 2B → 8B improves most on long-horizon reasoning.
• Token budget sweep (γ=2): SR 70.8 → 74.6 from B=2048 → 4608; OSR peaks at B=3584 then dips — more context helps until allocation quality binds.
• γ sweep (B=3072): OSR climbs monotonically to γ=3.5 (82.6) while SR peaks at γ=3.0 — recency bias helps find goals; too much of it erodes strict success and path quality. Together the sweeps confirm the interface parameters are live controls with task-legible trade-offs, not decoration.

Unitree Go2 quadruped, two serving modes: remote server 196 ms avg (5.1 Hz, higher variance) vs on-device Jetson Thor with FP8 + TensorRT 204 ms (4.9 Hz, more stable — preferred for tracking). Demonstrations: 21.78 m pure-language VLN in an unseen exhibition hall including a reverse command that walks the robot backward along its entire route to the start pose; fine-grained apartment commands ("stop at the nightstand on the left side of the bed"); and a full agentic episode ("check whether a green umbrella was left at Cotti Coffee") exercising the planner–executor–notebook loop end to end.

Three program-level observations:

1. The doctrine bends per domain. Language-as-interface and co-training-as-anti-collapse carry over intact; backbone choice, action decoder, and λ do not. Navigation's action space is low-dimensional enough that an MLP under MSE suffices — indirect support for VLM4VLA's minimal-harness methodology being production-viable in the right domain.
2. It supersedes Qwen-VLA's navigation claim. The May generalist argued one model should do manipulation + navigation; five weeks later the suite's navigation specialist beats it by 15+ pp on the same benchmarks. The suite decomposition is empirically vindicated on this axis.
3. The System-2 slot is now demonstrated, not just advertised. Review-Qwen-Team-VLA §6 flagged the agentic planner as the suite's unshipped keystone — the EQA results and the real-robot Cotti-Coffee episode show it working (as a system built around Qwen3.6-Plus, though not as a released product).

1. Token allocation is a heuristic — the authors say so explicitly and invite a principled replacement.
2. Tracking SR trails specialists (77–79% vs 86–87%) despite the best TR — the multi-task trade-off is real, not hidden.
3. Zero-shot closed-loop driving is weak (AlpaSim 0.15–0.17 vs 0.72): NAVSIM's 91.4 PDMS depends on prompt-injected GT ego-history (−11.9 without it) and does not transfer to at-fault closed-loop rollouts.
4. VLN-PE fall rate is ~10× baselines (3.83–4.05% vs 0.22–0.45%): waypoints that a physics-simulated body cannot always follow safely — the gap between waypoint prediction and full-kinematics locomotion is unresolved.

1. No weights (suite README: no release plan) — every number, including the RoboChallenge-style EQA leaderboard claims, is externally unreproducible; the planner (Qwen3.6-Plus) is also a closed model, so the agentic system is doubly irreproducible.
2. Configuration selection is unlearned. The planner picks Φ by LLM judgment with platform defaults; there is no ablation of how much planner configuration-switching contributes to the EQA gains vs a fixed sensible Φ — the headline agentic claim lacks its own component ablation.
3. Benchmark-adjacent training data. Instruction-following data is built from R2R/RxR training splits (5.63M samples from ~30K clips via augmentation) and tracking data from EVT-Bench's own training distribution; evaluation is on val-unseen/test splits as standard, but the massive per-benchmark augmentation ratio makes cross-model data-scale comparisons (e.g. vs NavFoM) hard to interpret as architecture wins alone.
4. Waypoint-only interface. K=8 (x,y,θ) waypoints presume a downstream locomotion controller; stairs, doors requiring interaction, and manipulation-coupled navigation (open the door, then go through) are outside the interface — exactly the seam where RobotManip and RobotNav would need to compose, and no cross-suite composition is demonstrated.
5. Embodiment preamble is binary in practice ("robot" vs "car"); the claimed extensibility to drones/quadrupeds-as-distinct-embodiments is asserted via the Go2 demo but not systematically evaluated across platforms.

• arXiv: https://arxiv.org/abs/2606.18112 · PDF: https://arxiv.org/pdf/2606.18112
• Blog: https://qwen.ai/blog?id=qwen-robotnav
• Code (docs only): https://github.com/QwenLM/Qwen-RobotNav

• Qwen Team's VLA Program — the cross-paper program review
• Qwen-RobotManip · Qwen-RobotWorld — the suite siblings
• Qwen-VLA — the generalist whose navigation numbers this paper supersedes
• NavFoM — the prior unified-navigation foundation model (same first-author lineage)
• CompassNav · CE-Nav — other 2026 navigation entries in this wiki
• VLM4VLA — the minimal-harness methodology this model unknowingly productionizes

← Back to Home