Finding information in hour-long videos is a challenging task even for top-performing Vision Language Models (VLMs), as encoding visual content quickly exceeds available context windows. To tackle this challenge, we present FALCONEye1, a novel video agent based on a training-free, model-agnostic meta-architecture composed of a VLM and a Large Language Model (LLM). FALCONEye answers open-ended questions using an exploration-based search algorithm guided by calibrated confidence from the VLM’s answers. We also introduce the FALCON-Bench benchmark, extending Question Answering problem to Video Answer Search—requiring models to return both the answer and its supporting temporal window for open-ended questions in hour-long videos. With just a 7B VLM and a lightweight LLM, FALCONEye outscores all open-source 7B VLMs and comparable agents in FALCON-Bench. It further demonstrates its generalization capability in MLVU benchmark with shorter videos and different tasks, surpassing GPT-4o on single-detail tasks while slashing inference cost by roughly an order of magnitude.

Sleep is fundamental to health, and society is more and more aware of the impact and relevance of sleep disorders. Traditional diagnostic methods, like polysomnography, are intrusive and resource-intensive. Instead, research is focusing on developing novel, less intrusive or portable methods that combine intelligent sensors with activity recognition for diagnosis support and scoring. Event cameras offer a promising alternative for automated, in-home sleep activity recognition due to their excellent low-light performance and low power consumption. This work introduces EventSleep2-data , a significant extension to the EventSleep dataset, featuring 10 complete night recordings (around 7 h each) of volunteers sleeping in their homes. Unlike the original short and controlled recordings, this new dataset captures natural, full-night sleep sessions under realistic conditions. This new data incorporates challenging real-world scene variations, an efficient movement-triggered sparse data recording pipeline, and synchronized 2-channel EEG data for a subset of recordings. We also present EventSleep2-net , a novel event-based sleep activity recognition approach with a dual-head architecture to simultaneously analyze motion classes and static poses. The model is specifically designed to handle the motion-triggered, sparse nature of complete night recordings. Unlike the original EventSleep architecture, EventSleep2-net can predict both movement and static poses even during long periods with no events. We demonstrate state-of-the-art performance on both EventSleep1-data, the original dataset, and EventSleep2-data, with comprehensive ablation studies validating our design decisions. Together, EventSleep2-data and EventSleep2-net overcome the limitations of the previous setup and enable continuous, full-night analysis for real-world sleep monitoring, significantly advancing the potential of event-based vision for sleep disorder studies.

Video is a crucial perception component in both robotics and wearable devices, two key technologies to enable innovative assistive applications, such as navigation and procedure execution assistance tools. Video understanding tasks are essential to enable these systems to interpret and execute complex instructions in real-world environments. One such task is step grounding, which involves identifying the temporal boundaries of activities based on natural language descriptions in long, untrimmed videos. This paper introduces Bayesian-VSLNet, a probabilistic formulation of step grounding that predicts a likelihood distribution over segments and refines it through Bayesian inference with temporal-order priors. These priors disambiguate cyclic and repeated actions that frequently appear in procedural tasks, enabling precise step localization in long videos. Our evaluations demonstrate superior performance over existing methods, achieving state-of-the-art results in the Ego4D Goal-Step dataset, winning the Goal Step challenge at the EgoVis 2024 CVPR. Furthermore, experiments on additional benchmarks confirm the generality of our approach beyond Ego4D. In addition, we present qualitative results in a real-world robotics scenario, illustrating the potential of this task to improve human–robot interaction in practical applications. Code is released at https://github.com/cplou99/BayesianVSLNet .

Short-Term object-interaction Anticipation (STA) consists in detecting the location of the next-active objects, the noun and verb categories of the interaction, as well as the time to contact from the observation of egocentric video. This ability is fundamental for wearable assistants to understand user's goals and provide timely assistance, or to enable human-robot interaction. In this work, we present a method to improve the performance of STA predictions. Our contributions are two-fold: 1) We propose STAformer and STAformer++, two novel attention-based architectures integrating frame-guided temporal pooling, dual image-video attention, and multiscale feature fusion to support STA predictions from an image-input video pair; 2) We introduce two novel modules to ground STA predictions on human behavior by modeling affordances. First, we integrate an environment affordance model which acts as a persistent memory of interactions that can take place in a given physical scene. We explore how to integrate environment affordances via simple late fusion and with an approach which adaptively learns how to best fuse affordances with end-to-end predictions. Second, we predict interaction hotspots from the observation of hands and object trajectories, increasing confidence in STA predictions localized around the hotspot. Our results show significant improvements on Overall Top-5 mAP, with gain up to $+23\%$+23% on Ego4D and $+31\%$+31% on a novel set of curated EPIC-Kitchens STA labels. We released the https://github.com/lmur98/AFFttention code, annotations, and pre-extracted affordances on Ego4D and EPIC-Kitchens to encourage future research in this area.

Environment understanding in egocentric videos is an important step for applications like robotics, augmented reality and assistive technologies. These videos are characterized by dynamic interactions and a strong dependence on the wearer’s engagement with the environment. Traditional approaches often focus on isolated clips or fail to integrate rich semantic and geometric information, limiting scene comprehension. We introduce Dynamic Image-Video Feature Fields (DIV-FF), a framework that decomposes the egocentric scene into persistent, dynamic, and actor-based components while integrating both image and video-language features. Our model enables detailed segmentation, captures affordances, understands the surroundings and maintains consistent understanding over time. DIV-FF outperforms state-of-the-art methods, particularly in dynamically evolving scenarios, demonstrating its potential to advance long-term, spatiotemporal scene understanding.

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) an Ego$\leftrightarrow$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, obtaining relative gains of +22% and +76% in the Ego2Exo and Exo2Ego IoU against the official challenge baselines, and a +13% and +6% compared with the SOTA with 1% of the training parameters.

Accurate affordance detection and segmentation with pixel precision is an important piece in many complex systems based on interactions, such as robots and assitive devices. We present a new approach to affordance perception which enables accurate multi-label segmentation. Our approach can be used to automatically extract grounded affordances from first person videos of interactions using a 3D map of the environment providing pixel level precision for the affordance location. We use this method to build the largest and most complete dataset on affordances based on the EPIC-Kitchen dataset, EPIC-Aff, which provides interaction-grounded, multi-label, metric and spatial affordance annotations. Then, we propose a new approach to affordance segmentation based on multi-label detection which enables multiple affordances to co-exists in the same space, for example if they are associated with the same object. We present several strategies of multi-label detection using several segmentation architectures. The experimental results highlight the importance of the multi-label detection. Finally, we show how our metric representation can be exploited for build a map of interaction hotspots in spatial action-centric zones and use that representation to perform a task-oriented navigation.

Single-view depth estimation can be remarkably effective if there is enough ground-truth depth data for supervised training. However, there are scenarios, especially in medicine in the case of endoscopies, where such data cannot be obtained. In such cases, multi-view self-supervision and synthetic-to-real transfer serve as alternative approaches, however, with a considerable performance reduction in comparison to supervised case. Instead, we propose a single-view self-supervised method that achieves a performance similar to the supervised case. In some medical devices, such as endoscopes, the camera and light sources are co-located at a small distance from the target surfaces. Thus, we can exploit that, for any given albedo and surface orientation, pixel brightness is inversely proportional to the square of the distance to the surface, providing a strong single-view self-supervisory signal. In our experiments, our self-supervised models deliver accuracies comparable to those of fully supervised ones, while being applicable without depth ground-truth data.

Affordances are a fundamental concept in robotics since they relate available actions for an agent depending on its sensory-motor capabilities and the environment. We present a novel Bayesian deep network to detect affordances in images, at the same time that we quantify the distribution of the aleatoric and epistemic variance at the spatial level. We adapt the Mask-RCNN architecture to learn a probabilistic representation using Monte Carlo dropout. Our results outperform the state-of-the-art of deterministic networks. We attribute this improvement to a better probabilistic feature space representation on the encoder and the Bayesian variability induced at the mask generation, which adapts better to the object contours. We also introduce the new Probability-based Mask Quality measure that reveals the semantic and spatial differences on a probabilistic instance segmentation model. We modify the existing Probabilistic Detection Quality metric by comparing the binary masks rather than the predicted bounding boxes, achieving a finer-grained evaluation of the probabilistic segmentation. We find aleatoric variance in the contours of the objects due to the camera noise, while epistemic variance appears in visual challenging pixels.

The integration of semantic information in a map allows robots to understand better their environment and make high-level decisions. In the last few years, neural networks have shown enormous progress in their perception capabilities. However, when fusing multiple observations from a neural network in a semantic map, its inherent overconfidence with unknown data gives too much weight to the outliers and decreases the robustness. To mitigate this issue we propose a novel robust fusion method to combine multiple Bayesian semantic predictions. Our method uses the uncertainty estimation provided by a Bayesian neural network to calibrate the way in which the measurements are fused. This is done by regularizing the observations to mitigate the problem of overconfident outlier predictions and using the epistemic uncertainty to weigh their influence in the fusion, resulting in a different formulation of the probability distributions. We validate our robust fusion strategy by performing experiments on photo-realistic simulated environments and real scenes. In both cases, we use a network trained on different data to expose the model to varying data distributions. The results show that considering the model's uncertainty and regularizing the probability distribution of the observations distribution results in a better semantic segmentation performance and more robustness to outliers, compared with other methods. Video - https://youtu.be/5xVGm7z9c-0

Bayesian optimization has become a popular method for applications, like the design of computer experiments or hyperparameter tuning of expensive models, where sample efficiency is mandatory. These situations or high-throughput computing, where distributed and scalable architectures are a necessity. However, Bayesian optimization is mostly sequential. Even parallel variants require certain computations between samples, limiting the parallelization bandwidth. Thompson sampling has been previously applied for distributed Bayesian optimization. But, when compared with other acquisition functions in the sequential setting, Thompson sampling is known to perform suboptimally. In this paper, we present a new method for fully distributed Bayesian optimization, which can be combined with any acquisition function. Our approach considers Bayesian optimization as a partially observable Markov decision process. In this context, stochastic policies, such as the Boltzmann policy, have some interesting properties which can also be studied for Bayesian optimization. Furthermore, the Boltzmann policy trivially allows a distributed Bayesian optimization implementation with high level of parallelism and scalability. We present results in several benchmarks and applications that shows the performance of our method.

In this paper, we tackle several problems that appear in robotics and autonomous systems: algorithm tuning, automatic control, and intelligent design. All those problems share in common that they can be mapped to global optimization problems where evaluations are expensive. Bayesian optimization (BO) has become a fundamental global optimization algorithm in many problems where sample efficiency is of paramount importance. BO uses a probabilistic surrogate model to learn the response function and reduce the number of samples required. Gaussian processes (GPs) have become a standard surrogate model for their flexibility to represent a distribution over functions. In a black-box settings, the common assumption is that the underlying function can be modeled with a stationary GP. In this paper, we present a novel kernel function specially designed for BO, that allows nonstationary behavior of the surrogate model in an adaptive local region. This kernel is able to reconstruct nonstationarity even with the irregular sampling distribution that arises from BO. Furthermore, in our experiments, we found that this new kernel results in an improved local search (exploitation), without penalizing the global search (exploration) in many applications. We provide extensive results in well-known optimization benchmarks, machine learning hyperparameter tuning, reinforcement learning, and control problems, and UAV wing optimization. The results show that the new method is able to outperform the state of the art in BO both in stationary and nonstationary problems.

Inference in the presence of outliers is an important field of research as outliers are ubiquitous and may arise across a variety of problems and domains. Bayesian optimization is method that heavily relies on probabilistic inference. This allows outstanding sample efficiency because the probabilistic machinery provides a memory of the whole optimization process. However, that virtue becomes a disadvantage when the memory is populated with outliers, inducing bias in the estimation. In this paper, we present an empirical evaluation of Bayesian optimization methods in the presence of outliers. The empirical evidence shows that Bayesian optimization with robust regression often produces suboptimal results. We then propose a new algorithm which combines robust regression (a Gaussian process with Student-t likelihood) with outlier diagnostics to classify data points as outliers or inliers. By using an scheduler for the classification of outliers, our method is more efficient and has better convergence over the standard robust regression. Furthermore, we show that even in controlled situations with no expected outliers, our method is able to produce better results.

Robust grasping is a major, and still unsolved, problem in robotics. Information about the 3D shape of an object can be obtained either from prior knowledge (e.g., accurate models of known objects or approximate models of familiar objects) or real-time sensing (e.g., partial point clouds of unknown objects) and can be used to identify good potential grasps. However, due to modeling and sensing inaccuracies, local exploration is often needed to refine such grasps and successfully apply them in the real world. The recently proposed unscented Bayesian optimization technique can make such exploration safer by selecting grasps that are robust to uncertainty in the input space (e.g., inaccuracies in the grasp execution). Extending our previous work on 2D optimization, in this paper we propose a 3D haptic exploration strategy that combines unscented Bayesian optimization with a novel collision penalty heuristic to find safe grasps in a very efficient way: while by augmenting the search-space to 3D we are able to find better grasps, the collision penalty heuristic allows us to do so without increasing the number of exploration steps.

Active policy search combines the trial-and-error methodology from policy search with Bayesian optimization to actively find the optimal policy. First, policy search is a type of reinforcement learning which has become very popular for robot control, for its ability to deal with complex continuous state and action spaces. Second, Bayesian optimization is a sample efficient global optimization method that uses a surrogate model, like a Gaussian process, and optimal decision making to carefully select each sample during the optimization process. Sample efficiency is of paramount importance when each trial involves the real robot, expensive Monte Carlo runs, or a complex simulator. Black-box Bayesian optimization generally assumes a cost function from a stationary process, because nonstationary modeling is usually based on prior knowledge. However, many control problems are inherently nonstationary due to their failure conditions, terminal states and other abrupt effects. In this paper, we present a kernel function specially designed for Bayesian optimization, that allows nonstationary modeling without prior knowledge, using an adaptive local region. The new kernel results in an improved local search (exploitation), without penalizing the global search (exploration), as shown experimentally in well-known optimization benchmarks and robot control scenarios. We finally show its potential for the design of the wing shape of a UAV.

Safe and robust grasping of unknown objects is a major challenge in robotics, which has no general solution yet. A promising approach relies on haptic exploration, where active optimization strategies can be employed to reduce the number of exploration trials. One critical problem is that certain optimal grasps discoverd by the optimization procedure may be very sensitive to small deviations of the parameters from their nominal values: we call these unsafe grasps because small errors during motor execution may turn optimal grasps into bad grasps. To reduce the risk of grasp failure, safe grasps should be favoured. Therefore, we propose a new algorithm, unscented Bayesian optimization, that performs efficient optimization while considering uncertainty in the input space, leading to the discovery of safe optima. The results highlight how our method outperforms the classical Bayesian optimization both in synthetic problems and in realistic robot grasp simulations, finding robust and safe grasps after a few exploration trials.

We present an active learning algorithm for the problem of body schema learning, i.e. estimating a kinematic model of a serial robot. The learning process is done online using Recursive Least Squares (RLS) estimation, which outperforms gradient methods usually applied in the literature. In addiction, the method provides the required information to apply an active learning algorithm to find the optimal set of robot configurations and observations to improve the learning process. By selecting the most informative observations, the proposed method minimizes the required amount of data. We have developed an efficient version of the active learning algorithm to select the points in real-time. The algorithms have been tested and compared using both simulated environments and a real humanoid robot.

Affordances are fundamental descriptors of relationships between actions, objects and effects. They provide the means whereby a robot can predict effects, recognize actions, select objects and plan its behavior according to desired goals. This paper approaches the problem of an embodied agent exploring the world and learning these affordances autonomously from its sensory experiences. Models exist for learning the structure and the parameters of a Bayesian Network encoding this knowledge. Although Bayesian Networks are capable of dealing with uncertainty and redundancy, previous work considered complete observability of the discrete sensory data, which may lead to hard errors in the presence of noise. In this paper we consider a probabilistic representation of the sensors by Gaussian Mixture Models (GMMs) and explicitly taking into account the probability distribution contained in each discrete affordance concept, which can lead to a more correct learning.

This paper presents a new particle method, with stochastic parameter estimation, to solve the SLAM problem. The underlying algorithm is rooted on a solid probabilistic foundation and is guaranteed to converge asymptotically, unlike many existing popular approaches. Moreover, it is efficient in storage and computation. The new algorithm carries out filtering only in the marginal filtering space, thereby allowing for the recursive computation of low variance estimates of the map. The paper provides mathematical arguments and empirical evidence to substantiate the fact that the new method represents an improvement over the existing particle filtering approaches for SLAM, which work on the joint path state space.

This paper addresses the consistency issue of the extended Kalman filter approach to the simultaneous localization and mapping (EKF-SLAM) problem. Linearization of the inherent nonlinearities of both the motion and the sensor models frequently drives the solution of the EKF-SLAM out of consistency specially in those situations where location uncertainty surpasses a certain threshold. This paper proposes a robocentric local map sequencing algorithm which: (a) bounds location uncertainty within each local map, (b) reduces the computational cost up to constant time in the majority of updates and (c) improves linearization accuracy by updating the map with sensor uncertainty level constraints. Simulation and large-scale outdoor experiments validate the proposed approach

This paper presents a robust algorithm for segmentation and line detection in 2D range scans. The described method exploits the multimodal probability density function of the residual error. It is capable of segmenting the range data in clusters, estimate the straight segments parameters, and estimate the scale of inliers error noise successfully, despite of high level of spurious data. No prior knowledge about the sensor and object properties is given to the algorithm. The mode seeking is based on mean shift algorithm, which has been widely used and tested in 3D laser scan segmentation, machine learning and pattern recognition applications. We show the reliability of the technique with experimental indoor and outdoor manmade environment. Compared with classical methods, a good compromise between false positive, false negative, wrong segment split and wrong segment merge is achieved, with improved accuracy in the estimated parameters.

This paper presents an experimentally validated alternative to the classical extended Kalman filter approach to the solution of the probabilistic state-space simultaneous localization and mapping (SLAM) problem. Several authors have reported the divergence of this classical approach due to the linearization of the inherent nonlinear nature of the SLAM problem. Hence, the approach described in this work aims to avoid the analytical linearization based on Taylor-series expansion of both the model and measurement equations by using the unscented filter. An innovation-based consistency checking validates the feasibility and applicability of the unscented SLAM approach to a real large-scale outdoor exploration mission.