Source-linked AI summary
PlatonicNav: Unveiling Semantic Correspondence in Navigation with Platonic Topological Maps
Junlin Long, Zeyu Zhang, Xu Deng, Yiran Wang, Yue Yang, Luke Borgnolo, Maxwell Twelftree, Yang Zhao
TL;DR
Existing navigation systems leave open whether independently trained vision and language encoders share semantic structure, and whether language goals can be grounded in purely vision-built maps without paired supervision. PlatonicNav tests this idea with a training-free Platonic Topological Map and blind matching, reporting cross-task, cross-modality, and cross-embodiment generalization across simulation benchmarks and Unitree Go2. The paper therefore presents ObjNav and VLN as interfaces to a shared object-centric semantic manifold, while identifying long instructions and modular components as current boundaries.
Problem
Prior VLN–ObjNav unification relies on architectural fusion, mixed-task training, or vision-language pretraining, while language grounding in object-centric maps still uses explicit cross-modal supervision.
Method
PlatonicNav uses a training-free Platonic Topological Map that fuses geometric and semantic distances from a self-supervised visual encoder and grounds language goals through blind matching without paired vision-language data.
Results
The framework shows cross-task, cross-modality, and cross-embodiment generalization on HM3D-IIN, OVON, R2R-CE on MP3D, and Unitree Go2 without explicit cross-modal training.
Takeaways & Limitations
ObjNav and VLN can be understood as different interfaces to a shared object-centric semantic manifold, supporting a representation-centric view of embodied navigation.
Takeaways & Limitations
Performance remains bounded by visual segmentation, blind matching, language-encoder expressiveness, and controller capability, with long natural-language instructions still challenging in R2R-CE.
Abstract
from arXiv · showhide
Embodied visual navigation, where an agent perceives a complex environment and acts to reach a goal from raw sensory input, underpins a wide range of applications such as household service robotics, assistive robotics, and large-scale autonomous exploration. However, recent attempts to unify vision-and-language navigation (VLN) and object goal navigation (ObjNav) remain at the level of architectural fusion, mixed-task training, and large vision-language pretraining, without examining whether independently trained vision and language encoders may already share a common semantic structure. Moreover, even object-centric topological maps still ground language goals through explicit cross-modal supervision such as CLIP or large vision-language models, leaving open whether such grounding is possible from a purely vision-built map. To address these challenges, we extend the Platonic Representation Hypothesis to embodied navigation and recast vision-only ObjNav, cross-modal ObjNav, and VLN as three different interfaces to the same object-centric semantic manifold. We further introduce PlatonicNav, a training-free framework whose Platonic Topological Map fuses geometric and semantic node distances from a self-supervised visual encoder, and grounds language goals via blind matching without any paired vision-language data. Extensive experiments on simulation benchmarks including HM3D-IIN, OVON, and R2R-CE on MP3D, together with deployment on Unitree Go2, demonstrate that PlatonicNav generalizes across tasks, modalities, and embodiments without explicit cross-modal training. Code: https://github.com/AIGeeksGroup/PlatonicNav. Website: https://aigeeksgroup.github.io/PlatonicNav.
1 Introduction
PlatonicNav argues that vision-only ObjNav, cross-modal ObjNav, and VLN can query a shared object-centric semantic structure despite independently trained encoders. It introduces a training-free map and blind-matching framework, evaluated across simulated and real embodiments without explicit cross-modal training.
- Existing VLN–ObjNav unification mainly uses architectural fusion, mixed-task training, and vision-language pretraining, leaving shared structure in independently trained encoders unresolved.
- The Platonic Representation Hypothesis predicts that vision-only ObjNav, cross-modal ObjNav, and VLN produce related trajectories by querying common semantic geometry through different goal interfaces.
- PlatonicNav grounds language goals in a vision-built Platonic Topological Map through blind matching, requiring no paired vision-language data.
- The framework formulates embodied navigation as a testable representation-level claim about real navigation trajectories across three task interfaces.
- Experiments on HM3D-IIN, OVON, R2R-CE on MP3D, and Unitree Go2 report cross-task, cross-modality, and cross-embodiment generalization without explicit cross-modal training.
- The paper frames navigation as operating over a shared object-centric semantic structure and organizes maps and policies around the geometry of meaning.
2 Related Work
Related work spans learning-based and foundation-model approaches to ObjNav, while language–vision binding is dominated by paired image-text supervision and joint multimodal representations.
- Object Goal Navigation evolved from imitation and reinforcement learning in Habitat toward semantic maps and zero-shot systems steered by foundation models.
- The prevailing language–vision paradigm uses massive paired image-text corpora to build joint embeddings, later lifted into open-vocabulary 3D representations.
3 Method
PlatonicNav represents navigation with a vision-built topological map whose nodes and edges encode object-centric semantic structure, then grounds language goals through relational blind matching. The framework connects vision-only ObjNav, cross-modal ObjNav, and VLN through shared semantic geometry and evaluates this premise with trajectory comparisons and cross-modal navigation results.
- Representation-level framing: PlatonicNav recasts vision-only ObjNav, cross-modal ObjNav, and VLN as goal interfaces over one object-centric semantic manifold.The interfaces differ in how the goal is presented, while navigation operates over shared semantic structure.
- Platonic Topological Map: The Platonic Topological Map represents image segments as object nodes and weights graph edges using geometric and visual-semantic distances.The map extends segment-based topological graphs by treating semantic proximity as part of node-to-node distance.
- Language-to-map grounding: Blind matching aligns visual clusters with language concepts by minimizing distortion between their pairwise relational distance structures.The resulting bijection assigns category labels without paired vision-language data, contrastive pretraining, or VLM supervision.
- Empirical motivation: The framework’s trajectory comparison tests whether vision-only ObjNav, VLN, and PlatonicNav follow related routes in matched scenes with matched targets.The comparison uses visited-node, semantic-scene-category, and landmark-intersection overlap rather than pixel-level path comparison; the analysis notes that trajectory overlap can also reflect environmental constraints and human-provided target correspondence.
- Empirical motivation: On HM3D-OVON, PlatonicNav’s non-trivial performance without explicit cross-modal training provides evidence that relational geometry between independently trained visual and language representations can support embodied goal grounding.The framework is also evaluated on HM3D-IIN, OVON, R2R-CE on MP3D, and a Unitree Go2 platform.
- Platonic Topological Map: Visual node embeddings are obtained from a self-supervised encoder by mean-pooling patch tokens within each segment mask and L2-normalizing the result.This representation makes Euclidean and cosine distances rank-equivalent and is intended to capture language-aligned semantic relationships without explicit visual-language supervision.
4 Experiments
PlatonicNav is evaluated across vision-only ObjNav, cross-modal ObjNav, VLN, and real-world quadruped deployment, with results indicating broad task and embodiment generalization.
- Simulation benchmarks: PlatonicNav achieves higher SPL and SSPL than ObjectReact on HM3D-IIN with goals grounded by ground-truth masks.This isolates the contribution of the Platonic Topological Map from cross-modal grounding differences.
- Simulation benchmarks: On HM3D-OVON, PlatonicNav outperforms the vast majority of cross-modal ObjNav methods on both SR and SPL.
- Benchmarks and metrics: The evaluation uses HM3D-IIN for instance navigation, HM3D-OVON for open-vocabulary category navigation, and R2R-CE for further independent-modality generalization.
- Simulation benchmarks: On R2R-CE Val-Unseen, PlatonicNav outperforms a considerable portion of VLN baselines, although some methods remain ahead.
- Real-world evaluation: Real-world evaluation deploys PlatonicNav on the Unitree Go2 quadruped using RGB observations and qualitative teach-and-repeat executions for ObjectNav and VLN.The real-world map is built offline from a human-guided teaching trajectory and reused during repeat execution.
5 Conclusion
PlatonicNav treats embodied navigation as traversal over a semantic manifold, using independently trained encoders and a calibrated hybrid topological map to support cross-modal grounding without paired data.
- PlatonicNav recasts vision-only ObjNav, cross-modal ObjNav, and VLN as interfaces to one object-centric semantic manifold.
- Blind matching aligns visual and language pairwise distance structures without paired data or contrastive training.
- Platonic Topological Maps augment segment-based graphs with relational semantic distances from a self-supervised visual encoder.
- The hybrid edge cost combines calibrated Platonic and geometric distances while preserving ObjectReact’s meter-scale navigation costs.Platonic distance is converted from dimensionless cosine distance to meter scale; λg = 1 recovers the original geometric graph.
C Top-down Trajectory Comparison
On HM3D-IIN, PlatonicNav and vision-only ObjNav produce generally similar trajectories, while PlatonicNav is more straightforward and usually more efficient.
- PlatonicNav’s trajectories are generally similar in shape to vision-only ObjNav trajectories but appear more straightforward.
- PlatonicNav acts more efficiently than vision-only ObjNav in most scenes.
D Real-world Implementation Platform
PlatonicNav was deployed on the quadruped Unitree Go2 and evaluated under object-goal and language-conditioned scenarios using onboard sensory inputs.
- PlatonicNav was deployed on a quadruped Unitree Go2 robot for real-world evaluation.
- The real-world protocol constructs topological maps from onboard sensory inputs and evaluates object-goal and language-conditioned navigation.The experiments test semantic-distance guidance under noisy observations and limited sensing conditions.
E Limitation and future work
PlatonicNav is an initial validation framework whose performance depends on several modular components, while long natural-language instructions remain challenging in R2R-CE.
- Limitation: PlatonicNav’s performance is bounded by visual segmentation, blind matching, language encoding, and ObjectReact-style controller quality.
- Limitation: R2R-CE results indicate that handling long natural-language instructions remains challenging.
- Future work: Future work targets robust vision-language matching, complex-instruction goal extraction, long-distance controller post-training, and open-vocabulary long-context navigation.
F.1 Selection of metric for Platonic distance
The metric study selects L2-normalized cosine distance for Platonic distance and examines semantic–geometric weighting and segmentation quality on HM3D-IIN.
- Metric selection: L2-normalized cosine distance outperforms L2-normalized euclidean distance for Platonic distance on HM3D-IIN.The comparison is reported using SPL and SSPL under the same configuration.
- Metric selection: L2-normalized cosine distance is chosen because it better captures semantic information from visual embedding space while accounting for embedding magnitude.This rationale accompanies the metric comparison and selection.
- Distance weighting: The Platonic Topological Map performs best with (λg, λs) = (0.8, 0.2) on HM3D-IIN.The authors interpret this configuration as evidence that semantic information improves navigation over purely geometric edge weights.
- Segmentation quality: Ground-truth segmentation yields relatively high performance for both ObjectReact and PlatonicNav, whereas FastSAM segmentation reduces SPL and SSPL to nearly half.The comparison evaluates how segmentation configuration affects both methods.
G.1 ObjectNav Qualitative Results
The ObjectNav qualitative results present teach- and repeat-phase visualizations across three tasks, documenting the paper’s qualitative evaluation examples.
- ObjectNav Task 1: ObjectNav Task 1 is illustrated in both teach and repeat phases.The referenced visualizations are Figures 7 and 8.
- Qualitative coverage: The qualitative examples cover ObjectNav task progression from Task 1 through Task 3.Each task includes paired teach- and repeat-phase figures.
- ObjectNav Task 2: ObjectNav Task 2 is illustrated in teach and repeat phases through qualitative visualizations.The referenced visualizations are Figures 9 and 10.
- ObjectNav Task 3: ObjectNav Task 3 is illustrated in teach and repeat phases through qualitative visualizations.The referenced visualizations are Figures 11 and 12.
H.1 VLN Simulation Results
The supplied VLN simulation materials comprise qualitative visualizations of navigation scenes, benchmark and software assets, and reported computational resources.
- VLN qualitative results: Additional VLN examples specify instructions such as exiting living rooms, reaching stairs, navigating hallways, and stopping near furniture or kitchen fixtures.The supplied passages identify concrete route descriptions and destinations.
- VLN qualitative results: VLN qualitative visualizations depict routes and stopping locations involving stairs, rugs, tables, chairs, kitchens, offices, and other indoor landmarks.The examples include teach and repeat phases as well as multiple simulation tasks.
- Simulation examples: The simulation materials include further visualizations for dining rooms, kitchens, fireplaces, offices, refrigerators, TV stands, chairs, sofas, and photos.These are listed across VLN and ObjNav simulation-task figures.
- Implementation assets: PlatonicNav uses listed implementations, checkpoints, datasets, and Habitat software with stated version or license information.The assets include DINOv3, SAM2, GTR-T5-base, HM3D-derived benchmarks, Matterport3D, and Habitat-Sim or Habitat-Lab.
- Computational resources: Four reported experiments consumed roughly 100–150 H100-hours, with comparable additional compute spent on earlier method iterations.Each run used one NVIDIA H100 GPU, while auxiliary mapping and feature-caching jobs used one NVIDIA L40 GPU.