Towards Social Navigation on a Quadruped Robot: Analyzing DINOv2 Visual Representations for Indoor Scene Understanding

Towards Social Navigation on a Quadruped Robot: Analyzing DINOv2 Visual Representations for Indoor Scene Understanding 1st Giovanna Guaragnella STIIMA-CNR Bari, Italy giov[email protected].it 2nd Annaclaudia Bono Politecnico di Bari & STIIMA CNR Bari, Italy [email protected].it 3rd Vito Ren` o STIIMA-CNR Bari, Italy [email protected].it Abstract—Social robots must perceive and navigate safely around humans in confined spaces. This work employs DINOv2 ViT-B/14 to extract visual features from RGB images captured by a quadruped robot in indoor environments by considering two input strategies (full-image and sliding-window), and by evaluating the model in configurations with and without token registers to assess their influence on feature extraction. Results demonstrate that the inclusion of registers, regardless of the input approach, improves spatial coherence, suggesting DINOv2 as a reliable backbone for future perception and navigation modules. Index Terms—Social robots, Artificial Intelligence, Navigation I. INTRODUCTION Robots designed to operate around humans must navigate safely, detect people, and understand complex environments in real time. Quadruped robots, such as the Unitree Go2, are becoming promising for social tasks, including elderly monitoring, fall detection, and people following, since they combine mobility with a user-friendly design [1], [2]. However, their onboard cameras observe the scene from a low and dynamic viewpoint, making perception challenging due to occlusions, quick viewpoint changes, and variable lighting. Traditional supervised pipelines often fail to generalize in complex and dynamic settings, while recent foundation models, such as DINOv2 [3], have demonstrated strong robustness in visual representation learning without task-specific training. Leveraging such models could support downstream tasks, including human tracking, semantic mapping, and safe navigation. This paper analyzes how the features extracted with the DINOv2 model could represent indoor environments from the onboard viewpoint of a robotic dog, as a first step toward foundationmodel-based perception for social navigation. II. MATERIALS AND METHODS The robotic platform adopted was Unitree Go2, equipped with an Intel RealSense RGB-D camera mounted at 45 cm from the ground, providing a low onboard viewpoint similar to a pet, enabling navigation in narrow or cluttered spaces. Visual data were collected in multiple small indoor environments (e.g., laboratories and under-table areas), under varying lighting conditions, simulating realistic scenarios for social assistance tasks. Fig. 1: Pipeline for feature extraction and feature map creation A feature extraction pipeline based on the DINOv2 ViTB/14 foundation model is implemented as shown in Figure 1. DINOv2 is a self-supervised vision transformer that processes an image as a sequence of 14×14 patches, each embedded into a 768-dimensional vector and refined through Transformer blocks with multi-head self-attention to produce contextual embeddings. To make these features interpretable and comparable, all 768-D patch embeddings are projected into a common 3D PCA space (first three components), mapped to RGB channels, and upsampled to image resolution by nearest-neighbor replication, preserving patch boundaries and producing consistent, visually interpretable feature maps across the dataset. III. EXPERIMENTS AND RESULTS All experiments were conducted on a workstation with an AMD EPYC 9654 CPU, 128 GB RAM, and two NVIDIA H100 GPUs (96 GB each) with the DINOv2 ViT-B/14 model in two variants: with and without registers. The registerenhanced version incorporates additional learnable tokens that aggregate global contextual information during training, potentially improving the semantic quality of the extracted 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629654 71 patch features. To evaluate the behaviour on robot-view data, two strategies were adopted to precess the input data (Table I). Both methods process the same 1280×720 images and produce patch-level embeddings arranged in a 768×52×92 feature map: •Full Image:the entire image is padded to dimensions divisible by the ViT patch size (14 pixels), and processed in a single forward pass through the model, which internally splits it into a 52×92 grid of non-overlapping patches. •Sliding Window:after padding, the image is scanned using overlapping 518×518 windows (each covering a 37×37 patch region, with a stride of 112 pixels). Each window is independently processed by the model, which internally splits it into 14×14 patches to extract local features. The resulting patch features are then projected back into their original positions and combined into a global 52×92 patch grid by weighted averaging of overlapping contributions. Qualitatively, as shown in Figure 2, the full-image approach produces smoother and more spatially coherent feature maps, effectively preserving the global structure of the scene, while the sliding-window method captures finer local details and object contours but introduces some visible discontinuities at window edges, caused by overlapping windows being processed independently and then merged. Feature maps generated without register tokens often lose local coherence, whereas those produced with registers maintain both local details and global context. Quantitatively, the variance explained by the first three principal components was also measured. Hence, for each 768dimensional feature set, PCA was applied, and the explained variance ratios of PC1–PC3 were compared between the models with and without token registers. Results (Figure 3) show an increase in the variance captured by PC1 and PC2 when register tokens are used, suggesting that meaningful information was strongly captured by the model, either in lower components. TABLE I: Description of the experiments, outlining the processing steps from the input image to the final feature map generation. Step Full Image Sliding Window Input 1280×720 1280×720 Padding 1288×728 1288×728 External windowing — 518×518 windows, stride 112 px (8 patch, ∼78% overlap) DINOv2 patching 52×92 (14×14) 37×37 (14×14) per window Feature extraction 1 forward pass ∼100 forward passes (518×518 windows) Aggregation — (direct) weighted averaging on 52×92 grid Final feature map 768×52×92 768×52×92 PCA global Incremental PCA (3D) global Incremental PCA (3D) Output image 1280×720 1280×720 Fig. 2: Feature maps produced by the full image (top) and sliding window (bottom) methods showing the original image (left) and the maps with (center) and without registers (right). Fig. 3: Explained variance (PC1–PC3) comparison between DINOv2 models with and without registers. IV. CONCLUSIONS This work investigates the feasibility of employing DINOv2 as a visual backbone for a quadruped robot operating in social robotics contexts. The results indicate that DINOv2 can extract coherent and stable visual features from the onboard perspective of a quadruped platform, even without task-specific training. These findings support the idea that foundation models can serve as versatile and general-purpose backbones for onboard perception in social robotics. Building on this, future efforts will focus on integrating these representations into real-time perception pipelines on Argo, enabling downstream modules for semantic mapping, dynamic obstacle detection, and human-aware navigation. Ultimately, this study represents an initial step toward developing visual perception modules for applications such as elderly monitoring, fall detection, and socially aware navigation in confined spaces. REFERENCES [1] P. Asgharian, A. Panchea, and F. Ferland, “A review on the use of mobile service robots in elderly care,” Robotics, vol. 11, no. 6, p. 127, 2022. doi: 10.3390/robotics11060127. [2] A. Elwaly, A. Abdellatif, and Y. El-Shaer, “New eldercare robot with path-planning and fall-detection capabilities,” Applied Sciences, vol. 14, no. 6, p. 2374, 2024. doi: 10.3390/app14062374. [3] M. Oquab, T. Darcet, T. Moutakanni, H. Vo, et al. “DINOv2: Learning Robust Visual Features without Supervision,” arXiv preprint arXiv:2304.07193, 2023. 72