To operate in the physical world, embodied agents must perceive their environment in an "always-on" fashion, selectively accessing the most informative sensors to balance energy constraints and task accuracy. Despite its importance for resource-constrained devices, energy-aware perception remains under-explored, with most prior work assuming unlimited compute. To address this, we introduce Ego-METAS: the first Egocentric online Multimodal Energy-efficient Temporal Action Segmentation benchmark. Ego-METAS provides a unified testbed of more than 100 hours of untrimmed egocentric video from EgoExo4D, CMU-MMAC, and CaptainCook4D, spanning 5 modalities (RGB, audio, gaze, IMU, and monochrome camera). We formulate an online temporal action segmentation task where models must dynamically select which sensors to activate at each timestep while strictly adhering to hardware-representative energy budgets. Alongside the benchmark, we release unified splits, cleaned annotations, pre-extracted features, and a diverse suite of baseline routing policies. Our evaluations show that optimal routing is highly scenario-dependent, and that existing policy-learning methods—designed primarily for trimmed clips—struggle to adapt to continuous, untrimmed environments. However, even simple dynamic fusion of complementary modalities (e.g., via random routing) proves critical for balancing predictive accuracy against strict energy budgets. Ultimately, Ego-METAS provides a standardized foundation to develop robust, cost-aware policies for autonomous, always-on embodied AI.

Existing methods for egocentric action recognition often rely solely on RGB videos, although additional modalities, e.g., audio, can improve accuracy in challenging scenarios. However, most multimodal approaches assume all modalities are available at inference, leading to significant accuracy drops, or even failure, when inputs are missing. To address this, we introduce KARMMA, a multimodal Knowledge distillation approach for egocentric Action Recognition robust to Missing ModAlities that requires no modality alignment across all samples during training or inference. KARMMA distills knowledge from a multimodal teacher into a multimodal student that benefits from all available modalities while remaining robust to missing ones, making it suitable for diverse scenarios without retraining. Our student uses approximately 50% fewer computational resources than our teacher, resulting in a lightweight and fast model. Experiments on Epic-Kitchens and Something-Something show that our student achieves competitive accuracy while significantly reducing accuracy drops under missing modality conditions.

Understanding the world from multiple perspectives is essential for intelligent systems operating together, where segmenting common objects across different views remains an open problem. We introduce a new approach that re-defines cross-image segmentation by treating it as a mask matching task. Our method consists of: (1) A Mask-Context Encoder that pools dense DINOv2 semantic features to obtain discriminative object-level representations from FastSAM mask candidates, (2) a Ego↔Exo Cross-Attention that fuses multi-perspective observations, (3) a Mask Matching contrastive loss that aligns cross-view features in a shared latent space and, (4) a Hard Negative Adjacent Mining strategy to encourage the model to better differentiate between nearby objects. O-MaMa achieves the state of the art in the Ego-Exo4D Correspondences benchmark.

One of the main challenges of visual prostheses is to augment the perceived information to improve the experience of its wearers. Given the limited access to implanted patients, in order to facilitate the experimentation of new techniques, this is often evaluated via Simulated Prosthetic Vision (SPV) with sighted people. In this work, we introduce a novel SPV framework and implementation that presents major advantages with respect to previous approaches. First, it is integrated into a robotics framework, which allows us to benefit from a wide range of methods and algorithms from the field (e.g. object recognition, obstacle avoidance, autonomous navigation, deep learning). Second, we go beyond traditional image processing with 3D point clouds processing using an RGB-D camera, allowing us to robustly detect the floor, obstacles and the structure of the scene. Third, it works either with a real camera or in a virtual environment, which gives us endless possibilities for immersive experimentation through a head-mounted display. Fourth, we incorporate a validated temporal phosphene model that replicates time effects into the generation of visual stimuli. Finally, we have proposed, developed and tested several applications within this framework, such as avoiding moving obstacles, providing a general understanding of the scene, staircase detection, helping the subject to navigate an unfamiliar space, and object and person detection. We provide experimental results in real and virtual environments.

Convolution kernels are the basic structural component of convolutional neural networks (CNNs). In the last years there has been a growing interest in fisheye cameras for many applications. However, the radially symmetric projection model of these cameras produces high distortions that affect the performance of CNNs, especially when the field of view is very large. In this work, we tackle this problem by proposing a method that leverages the calibration of cameras to deform the convolution kernel accordingly and adapt to the distortion. That way, the receptive field of the convolution is similar to standard convolutions in perspective images, allowing us to take advantage of pre-trained networks in large perspective datasets. We show how, with just a brief fine-tuning stage in a small dataset, we improve the performance of the network for the calibrated fisheye with respect to standard convolutions in depth estimation and semantic segmentation.