R2R-CE-VP: Visual Prompt Navigation Benchmark

• The paper introduces R2R-CE-VP, a novel benchmark that transforms language-guided navigation into a visual prompt-based task in continuous 3D environments.
• It details a four-step pipeline for generating allocentric top-view route maps from Matterport3D scenes, ensuring quality through rigorous filtering.
• The VPNet architecture is adapted for continuous settings using a ViT-based visual prompt encoder and graph-aware waypoint prediction to guide navigation.

R2R-CE-VP is a continuous-environment visual prompt navigation benchmark introduced within the Visual Prompt Navigation (VPN) framework. It is constructed by transforming R2R-CE episodes from language-guided continuous VLN into episodes in which the natural-language instruction is removed and replaced by a top-view visual prompt map that depicts the target route. The benchmark is implemented in the Habitat simulator and requires an agent to navigate in a continuous 3D indoor environment using first-person visual observations together with an allocentric, route-marked 2D map rather than text (Feng et al., 3 Aug 2025).

1. Definition, provenance, and scope

R2R-CE-VP is the continuous-environment counterpart of VPN and is built directly from the existing R2R-CE benchmark. The task inherits the continuous Habitat setting of R2R-CE but changes the instruction modality: the agent is no longer conditioned on natural language, and instead receives a pre-supplied top-view map with the intended trajectory visually annotated. The paper describes this as establishing “VPN in continuous environments (VPN-CE)” over Habitat, with a pretrained waypoint predictor used to align the formulation with VPN in discrete environments (Feng et al., 3 Aug 2025).

Within the supplied literature, this exact label is specific to embodied navigation. It should therefore be distinguished from unrelated uses of R2R in roll-to-roll manufacturing, including corona-enabled electrostatic printing for soft electronics (Wang et al., 2021) and multi-section R2R tension control with curriculum-based SAC (Li et al., 27 Feb 2026). Those works use the R2R abbreviation in manufacturing and control contexts and do not define the R2R-CE-VP benchmark.

A common misconception is to treat R2R-CE-VP as a minor visualization variant of R2R-CE. The task definition is more consequential than that. It replaces the linguistic instruction channel with a spatially grounded visual route specification, which changes the supervision format, the ambiguity profile of the task, and the encoder design required by the policy. Another misconception is to read “continuous” as direct low-level control over primitive motor commands in the policy network. In the benchmark formulation, the environment is continuous, but high-level policy decisions are made over predicted navigable waypoints plus a dedicated stop action.

2. Dataset construction and visual prompt generation

R2R-CE-VP is constructed by transferring R2R-CE episodes into the VPN format through a four-step pipeline. First, for each episode, a 2D top-view map is generated for the corresponding scene from the 90 Matterport3D environments, following a procedure similar to LED. Second, the pixel coordinates of all viewpoints are computed on the top-view map. Third, those projected viewpoints are connected sequentially with arrows, and the map is center-cropped around the trajectory. The crop is defined as a square region centered on the trajectory whose side length equals the maximum of trajectory width and trajectory height plus 60 pixels. Fourth, the bounding box of all non-zero pixels is detected and used to remove unnecessary black borders (Feng et al., 3 Aug 2025).

The benchmark includes filtering. Some scenes or floors are excluded because of poor top-down reconstruction quality. For MP3D multi-floor scenes, floors with poor reconstruction are manually excluded. For HM3D data used in augmentation, only the first floor is retained. In the continuous setting, some R2R-CE viewpoints cannot be mapped to MP3D 3D meshes; only valid viewpoints are retained, and episodes containing invalid viewpoints are removed. This filtering is one reason the benchmark has slightly fewer scenes and episodes than the R2R-CE source set.

Dataset	Scenes	Episodes
R2R / R2R-CE source	61	3603
R2R-CE-VP	60	3597
PREVALENT-derived VPN continuous data	59	88,694

Multi-floor handling is addressed through Order-Aware Floor Concatenation (OAFC), which preserves floor traversal order by representing an episode with multiple prompt maps when the trajectory spans multiple floors. The reported MP3D floor statistics are 32 scenes with 1 floor, 32 with 2 floors, 17 with 3 floors, 8 with 4 floors, and 1 with 5 floors. All visual prompt maps are resized to $224 \times 224$ before encoding (Feng et al., 3 Aug 2025).

The prompt itself is an allocentric raster image. It is globally aligned to the scene layout, and its visual content primarily marks the navigation trajectory, key waypoints, and directional progression through arrows or lines. The exact projection equations, meters-per-pixel resolution, and origin conventions are not provided in the supplied text.

3. Task formulation in continuous environments

In R2R-CE-VP, the agent must navigate from a start location to a goal location in a continuous 3D indoor scene using only first-person visual observations and the visual prompt map. At each step, the model observes panoramic views

$V_t=\{v_{t,i}\}_{i=1}^{n}.$

For the continuous case, those observations are paired with a pretrained waypoint predictor that proposes navigable candidates, making the policy formulation analogous to graph-based navigation in discrete settings (Feng et al., 3 Aug 2025).

The action space is continuous at the environment level but waypoint-based at the policy level. After graph construction, the model outputs scores over candidate navigable nodes together with a stop action,

$$\hat{v}=(\text{stop}, nav_1, nav_2, \dots, nav_{v'}).$$

The navigation objective is to follow the prompted route and terminate sufficiently close to the goal. Success is defined under the standard R2R-CE protocol as final navigation error $< 3$ meters.

The prompt representation is not merely a scene map; it is a scene map with explicit route markup. The paper studies several prompt forms in ablation and reports that the best-performing type is “lines only,” namely route rendering on cropped top-view maps without extra text annotations. That ablation is reported for the discrete benchmark rather than specifically for R2R-CE-VP, but it is used to motivate the general visual prompt design. This suggests that, within VPN, the dominant instruction signal is the route geometry itself rather than textual tags or uncropped environmental context (Feng et al., 3 Aug 2025).

Conceptually, R2R-CE-VP occupies an intermediate position between VLN and ImageNav. It retains multi-step trajectory guidance, as in VLN, but encodes that guidance visually rather than linguistically. The prompt is allocentric and spatially explicit rather than relative to the agent’s current heading.

4. VPNet architecture and training regime

The continuous baseline for R2R-CE-VP is VPNet in continuous mode. The paper states explicitly that discrete VPNet is based on DUET, whereas continuous VPNet is based on ETPNav. In both cases, the original instruction encoder is replaced by a ViT-based encoder operating on top-view prompt maps. The architecture comprises a Visual Prompt Encoder, a Node Embedding module, and a Graph-aware Cross-modal Encoder; for continuous environments, only the graph-aware encoder is used (Feng et al., 3 Aug 2025).

Prompt encoding is formalized as

$\mathcal{P}_i^o = \text{ViT}(\mathcal{P}_i)+b_i,$

$$\mathcal{P} = [\mathcal{P}_1^o,\mathcal{P}_2^o, \dots,\mathcal{P}_k^o],$$

where $\mathcal{P}_i$ is the $i$-th prompt map, $b_i$ is an order embedding encoding floor traversal sequence, and $k$ is the number of prompt maps. The implementation uses ViT-B/16 pretrained on ImageNet both for panoramic observation features and for prompt-map encoding.

The policy incrementally builds a topological graph,

$V_t=\{v_{t,i}\}_{i=1}^{n}.$0

whose nodes include the current node, visited nodes, and navigable nodes. In continuous environments, navigable candidates are supplied by the pretrained waypoint predictor. Each node embedding combines step embeddings, position embeddings, and contextualized view features. Cross-modal fusion is performed by a graph transformer, after which a feed-forward network produces global action scores over node candidates and stop.

Training uses a weighted combination of behavior cloning and DAgger supervision, with $V_t=\{v_{t,i}\}_{i=1}^{n}.$1. The pseudo-label for DAgger is the next node on the shortest path to the destination over the partial graph built under on-policy sampling. The reported hidden size is 768; the visual prompt encoder has 12 layers, the panoramic view encoder 2 layers, the graph-aware cross-modal encoder 4 layers, and the local cross-modal encoder 4 layers, although the local encoder is not used in the continuous setting (Feng et al., 3 Aug 2025).

Two augmentation mechanisms are central to the training regime. View-level augmentation includes prompt-view augmentation, which rotates top-view prompt maps by $V_t=\{v_{t,i}\}_{i=1}^{n}.$2, and agent-view augmentation, which samples the initial heading from $V_t=\{v_{t,i}\}_{i=1}^{n}.$3. Trajectory-level augmentation adds PREVALENT-derived continuous trajectories; in the reported continuous results, the augmentation setting is effectively R2R plus PREVALENT-derived data rather than ScaleVLN.

5. Evaluation protocol and reported performance

R2R-CE-VP uses the same metric family as VLN. The reported continuous metrics are TL, NE, SR, OSR, and SPL, although the main continuous results table reports NE, SR, and SPL. Success is defined by the threshold $V_t=\{v_{t,i}\}_{i=1}^{n}.$4 meters, and oracle success assumes oracle stopping. Evaluation is conducted online in Habitat continuous 3D indoor environments (Feng et al., 3 Aug 2025).

The main reported results for R2R-CE-VP are as follows.

Training setup	Val seen	Val unseen
VPNet (R2R)	NE 4.70; SR 57.75; SPL 53.58	NE 5.96; SR 42.09; SPL 36.46
VPNet (R2R+PRE)	NE 2.27; SR 84.88; SPL 80.00	NE 5.90; SR 49.27; SPL 42.12

These numbers indicate that PREVALENT-derived trajectory-level augmentation produces a large gain on val seen and a smaller but still meaningful improvement on val unseen. The strongest reported configuration in the supplied main results is therefore VPNet trained with R2R plus PREVALENT-derived data.

The paper also juxtaposes R2R-CE-VP with language-guided VLN-CE baselines evaluated on R2R-CE, including Seq2Seq, WPN, CMA, GridMM, ETPNav, and NaVILA. On val unseen, those listed language-guided baselines span SR values from 32 to 57 and SPL values from 30 to 49, while VPNet on R2R-CE-VP reaches SR 49.27 and SPL 42.12 with R2R+PRE training. The authors use this comparison to argue that visual prompts are competitive and viable, but the paper also makes clear that this is not same-task evaluation because the instruction modality differs.

A recurrent interpretive issue concerns prompt design. The prompt-type ablation reported for the discrete benchmark shows that “lines only” on cropped maps performs best, with val unseen SR 65.92 and SPL 56.17, whereas a full uncropped map achieves SR 33.94 and SPL 29.36. Although those numbers are not reported specifically for R2R-CE-VP, they support the broader claim that VPN performance depends strongly on prompt-form design rather than on the mere presence of a map.

6. Position within continuous-navigation research

R2R-CE-VP is best understood as a benchmark for replacing natural-language instruction with an allocentric, spatially grounded visual route specification. In that sense, it is not simply a derivative of R2R-CE, but a change in the instruction interface for continuous-environment navigation. The benchmark retains Habitat, panoramic observations, waypoint prediction, and topological planning machinery while altering how route information is presented to the policy (Feng et al., 3 Aug 2025).

Its closest methodological neighbors are graph-based continuous VLN systems that already plan over waypoint or node candidates instead of raw primitive actions. A later example is ETP-R1, which targets R2R-CE and RxR-CE rather than R2R-CE-VP, but shares the core structure of online topological map construction, a pretrained waypoint predictor, graph-node selection, and deterministic low-level execution in continuous Habitat environments. ETP-R1 reports $V_t=\{v_{t,i}\}_{i=1}^{n}.$5 SR and $V_t=\{v_{t,i}\}_{i=1}^{n}.$6 SPL on R2R-CE val unseen and is therefore relevant as an adjacent planning substrate rather than as a direct result on the R2R-CE-VP benchmark (Ye et al., 24 Dec 2025).

Several misconceptions are therefore worth separating. R2R-CE-VP is not a language benchmark with an auxiliary map overlay; it is a visual-prompt benchmark. It is not equivalent to image-goal navigation, because the prompt specifies a multi-step route rather than only a terminal appearance target. It is also not a purely discrete graph task, because evaluation occurs in continuous Habitat environments with continuous path execution and continuous-distance success criteria. A plausible implication is that R2R-CE-VP serves as a testbed for studying whether route geometry alone, encoded as a top-view prompt, can substitute for linguistic trajectory descriptions without changing the continuous-environment execution substrate.

In the supplied literature, that is the benchmark’s central significance: it operationalizes a visual alternative to language-guided continuous VLN by converting R2R-CE episodes into top-view route-following problems, pairing allocentric prompt maps with first-person observations, and evaluating the resulting policies under the standard continuous navigation protocol (Feng et al., 3 Aug 2025).