Image goal navigation requires two different skills: firstly, core navigation skills, including the detection of free space and obstacles, and taking decisions based on an internal representation; and secondly, computing directional information by comparing visual observations to the goal image. Current state-of-the-art methods either rely on dedicated image-matching, or pre-training of computer vision modules on relative pose estimation. In this paper, we study whether this task can be efficiently solved with end-to-end training of full agents with RL, as has been claimed by recent work. A positive answer would have impact beyond Embodied AI and allow training of relative pose estimation from reward for navigation alone. In this large experimental study we investigate the effect of architectural choices like late fusion, channel stacking, space-to-depth projections and cross-attention, and their role in the emergence of relative pose estimators from navigation training. We show that the success of recent methods is influenced up to a certain extent by simulator settings, leading to shortcuts in simulation. However, we also show that these capabilities can be transferred to more realistic setting, up to some extent. We also find evidence for correlations between navigation performance and probed (emerging) relative pose estimation performance, an important sub skill.

Progress in Embodied AI has made it possible for end-to-end-trained agents to navigate in photo-realistic environments with high-level reasoning and zero-shot or language-conditioned behavior, but benchmarks are still dominated by simulation. In this work, we focus on the fine-grained behavior of fast-moving real robots and present a large-scale experimental study involving 262 navigation episodes in a real environment with a physical robot, where we analyze the type of reasoning emerging from end-to-end training. In particular, we study the presence of realistic dynamics which the agent learned for open-loop forecasting, and their interplay with sensing. We analyze the way the agent uses latent memory to hold elements of the scene structure and information gathered during exploration. We probe the planning capabilities of the agent, and find in its memory evidence for somewhat precise plans over a limited horizon. Furthermore, we show in a post-hoc analysis that the value function learned by the agent relates to long-term planning. Put together, our experiments paint a new picture on how using tools from computer vision and sequential decision making have led to new capabilities in robotics and control. An interactive tool is available [here].

One key aspect of spatially aware robots is the ability to "find their bearings", ie. to correctly situate themselves or previously seen spaces. In this work, we focus on this particular scenario of continuous robotics operations, where information observed before an actual episode start is exploited to optimize efficiency. We introduce a new model, "Kinaema" and agent, capable of integrating a stream of visual observations while moving in a potentially large scene, and upon request, processing a query image and predicting the relative position of the shown space with respect to its current position. Our model does not explicitly store an observation history, therefore does not have hard constraints on context length. It maintains an implicit latent memory, which is updated by a transformer in a recurrent way, compressing the history of sensor readings into a compact representation. We evaluate the impact of this model in a new downstream task we call "Mem-Nav", targeting continuous robotics operations. We show that our large-capacity recurrent model maintains a useful representation of the scene, navigates to goals observed before the actual episode start, and is computationally efficient, in particular compared to classical transformers with attention over an observation history.

As robots increasingly share spaces with people, it becomes important for them to behave according to our social norms. In this paper, we explore the problem of finding socially acceptable locations for a robot to wait for a shared elevator by learning from expert annotations. Access to relevant, unlabeled data is however scarce in this setting and annotations expensive to gather, as they require explicit knowledge about the social norms, the robot, and the service it carries out. We tackle this low-data regime as follows. First, we use Procedural Content Generation to generate plausible waiting scenes to be annotated. Second, we leverage available sociological studies and operationalize relevant social norms as feature maps. We train a variety of models with only 125 procedurally-generated expert-annotated scenes, testing the impact of the proposed feature maps. In our ablation study, the feature maps help the models’ performance and their generalization capabilities to non-synthetic, real scenes. We inspect the decisions taken by the best models, probing their strengths and weaknesses, and identifying general issues and discussing potential solutions.

Bird’s-eye view (BEV) maps are an important geometrically structured representation widely used in robotics, in particular self-driving vehicles and terrestrial robots. Existing algorithms either require depth information for the geometric projection, which is not always reliably available, or are trained end-to-end in a fully supervised way to map visual first-person observations to BEV representation, and are therefore restricted to the output modality they have been trained for. In contrast, we propose a new model capable of performing zero-shot projections of any modality available in a first person view to the corresponding BEV map. This is achieved by disentangling the geometric inverse perspective projection from the modality transformation, e.g. RGB to occupancy. The method is general and we showcase experiments projecting to BEV three different modalities: semantic segmentation, motion vectors and object bounding boxes detected in first person. We experimentally show that the model outperforms competing methods, in particular the widely used baseline resorting to monocular depth estimation.

In the context of autonomous navigation of terrestrial robots, the creation of realistic models for agent dynamics and sensing is a widespread habit in the robotics literature and in commercial applications, where they are used for model based control and/or for localization and mapping. The more recent Embodied AI literature, on the other hand, focuses on modular or end-to-end agents trained in simulators like Habitat or AI-Thor, where the emphasis is put on photo-realistic rendering and scene diversity, but high-fidelity robot motion is assigned a less privileged role. The resulting sim2real gap significantly impacts transfer of the trained models to real robotic platforms. In this work we explore end-to-end training of agents in simulation in settings which minimize the sim2real gap both, in sensing and in actuation. Our agent directly predicts (discretized) velocity commands, which are maintained through closed-loop control in the real robot. The behavior of the real robot (including the underlying low-level controller) is identified and simulated in a modified Habitat simulator. Noise models for odometry and localization further contribute in lowering the sim2real gap. We evaluate on real navigation scenarios, explore different localization and point goal calculation methods and report significant gains in performance and robustness compared to prior work.

Agents navigating in 3D environments require some form of memory, which should hold a compact and actionable representation of the history of observations useful for decision taking and planning. In most end-to-end learning approaches the representation is latent and usually does not have a clearly defined interpretation, whereas classical robotics addresses this with scene reconstruction resulting in some form of map, usually estimated with geometry and sensor models and/or learning. In this work we propose to learn an actionable representation of the scene independently of the targeted downstream task and without explicitly optimizing reconstruction. The learned representation is optimized by a blind auxiliary agent trained to navigate with it on multiple short sub episodes branching out from a waypoint and, most importantly, without any direct visual observation. We argue and show that the blindness property is important and forces the (trained) latent representation to be the only means for planning. With probing experiments we show that the learned representation optimizes navigability and not reconstruction. On downstream tasks we show that it is robust to changes in distribution, in particular the sim2real gap, which we evaluate with a real physical robot in a real office building, significantly improving performance.

Navigation of terrestrial robots is typically addressed either with localization and mapping (SLAM) followed by classical planning on the dynamically created maps, or by machine learning (ML), often through end-to-end training with reinforcement learning (RL) or imitation learning (IL). Recently, modular designs have achieved promising results, and hybrid algorithms that combine ML with classical planning have been proposed. Existing methods implement these combinations with handcrafted functions, which cannot fully exploit the complementary nature of the policies and the complex regularities between scene structure and planning performance. Our work builds on the hypothesis that the strengths and weaknesses of neural planners and classical planners follow some regularities, which can be learned from training data, in particular from interactions. This is grounded on the assumption that, both, trained planners and the mapping algorithms underlying classical planning are subject to failure cases depending on the semantics of the scene and that this dependence is learnable: for instance, certain areas, objects or scene structures can be reconstructed easier than others. We propose a hierarchical method composed of a high-level planner dynamically switching between a classical and a neural planner. We fully train all neural policies in simulation and evaluate the method in both simulation and real experiments with a LoCoBot robot, showing significant gains in performance, in particular in the real environment. We also qualitatively conjecture on the nature of data regularities exploited by the high-level planner.

Recently a series of key factors are deeply transforming lighting systems. Single lights can be controlled individually and with advanced rendering capabilities. Furthermore, a shift is underway from independent light sources to integrated lighting installations. These factors fuel the need for intuitive, flexible control techniques that can fully exploit the rendering capabilities of a lighting infrastructure. In this paper, we present a novel framework to automatically create lighting atmospheres in any type of environment. The desired light settings are derived by approximation of user-defined input images or videos. This input modality gives an immediate and intuitive representation of a lighting atmosphere and allows the user to translate into a light setting any visual content. The effectiveness and versatility of the proposed solution is demonstrated and discussed by deploying the system in several challenging, real-life application scenarios.

While it has been widely proved that breastfeeding is the healthiest feeding option for a baby and its mother, the mechanisms by which a baby removes milk from the breast are still not completely known. Partly this is due to the lack of tools to analyze images of the infant oral cavity during feeding automatically and quantitatively. In this paper we propose two methods for analyzing ultrasound videos to automatically detect relevant events such sucking and swallowing of milk and to discriminate different types of tongue action during milk removal. The proposed algorithms provide, for the first time, quantitative indications of the type of activities carried out during breastfeeding by the baby, promising unprecedented advancements in the field.

The advent of integrated lighting installations, consisting of individually controllable lamps with advanced rendering capabilities, is fundamentally transforming lighting. This brings a need for an intuitive control capable of fully exploiting the rendering capabilities of the complete lighting infrastructure. In this paper we present a new method to automatically create lighting atmospheres in any type of environment, that allows for an natural interaction with the lighting system and generates unique, suggestive effects. To prove the effectiveness and versatility of the proposed solution, we deploy the system in several application scenarios, and discuss the results.

Flower Power is an interactive installation that creates immersive light atmospheres by capturing the colors of flowers and re-projecting them in the environment. A camera is suspended over a flower bed and its position depends on the positions of the visitors around the installation, which are detected using presence sensors. The image captured by the camera is rendered into the surrounding environment as a suggestive, saturated light scene. The participants can thus give the power to their favorite flower to set the overall light experience. In this paper we describe the motivation and the realization of the installation, and we discuss insights and reactions collected during the 2010 STRP Festival.

In this paper, we propose a novel method which is able to detect and separate audiovisual sources present in a scene. Our method exploits the correlation between the video signal captured with a camera and a synchronously recorded one-microphone audio track. In a first stage, audio and video modalities are decomposed into relevant basic structures using redundant representations. Next, synchrony between relevant events in audio and video modalities is quantified. Based on this co-occurrence measure, audiovisual sources are counted and located in the image using a robust clustering algorithm that groups video structures exhibiting strong correlations with the audio. Next periods where each source is active alone are determined and used to buildspectralGaussian mixture models (GMMs) characterizing the sources acoustic behavior. Finally, these models are used to separate the audio signal in periods during which several sources are mixed. The proposed approach has been extensively tested on synthetic and natural sequences composed of speakers and music instruments. Results show that the proposed method is able to successfully detect, localize, separate, and reconstruct present audiovisual sources.

In recent years, works on geometric multidimensional signal representations have established a close relation with signal expansions on redundant dictionaries. For this purpose, matching pursuits (MP) have shown to be an interesting tool. Recently, most important limitations of MP have been underlined, and alternative algorithms like weighted-MP have been proposed. This work explores the use of weighted-MP as a new framework for motion-adaptive geometric video approximations. We study a novel algorithm to decompose video sequences in terms of few, salient video components that jointly represent the geometric and motion content of a scene. Experimental coding results on highly geometric content reflect how the proposed paradigm exploits spatio-temporal video geometry. Two-dimensional weighted-MP improves the representation compared to those based on 2-D MP. Furthermore, the extracted video components represent relevant visual structures with high saliency. In an example application, such components are effectively used as video descriptors for the joint audio-video analysis of multimedia sequences.

A novel model is presented to learn bimodally informative structures from audio-visual signals. The signal is represented as a sparse sum of audio-visual kernels. Each kernel is a bimodal function consisting of synchronous snippets of an audio waveform and a spatio-temporal visual basis function. To represent an audio-visual signal, the kernels can be positioned independently and arbitrarily in space and time. The proposed algorithm uses unsupervised learning to form dictionaries of bimodal kernels from audio-visual material. The basis functions that emerge during learning capture salient audio-visual data structures. In addition, it is demonstrated that the learned dictionary can be used to locate sources of sound in the movie frame. Specifically, in sequences containing two speakers, the algorithm can robustly localize a speaker even in the presence of severe acoustic and visual distracters.

In this work we present a method to perform a complete audiovisual source separation without need of previous information. This method is based on the assumption that sounds are caused by moving structures. Thus, an efficient representation of audio and video sequences allows to build relationships between synchronous structures on both modalities. A robust clustering algorithm groups video structures exhibiting strong correlations with the audio so that sources are counted and located in the image. Using such information and exploiting audio-video correlation, the audio sources activity is determined. Next, spectral Gaussian Mixture Models (GMMs) are learnt in time slots with only one source active so that it is possible to separate them in case of an audio mixture. Audio source separation performances are rigorously evaluated, clearly showing that the proposed algorithm performs efficiently and robustly.

We present a general framework and an efficient algorithm for tracking relevant video structures. The structures to be tracked are implicitly defined by a matching pursuit procedure that extracts and ranks the most important image contours. Based on the ranking, the contours are automatically selected to initialize a particle filtering tracker. The proposed algorithm deals with salient video entities whose behavior has an intuitive meaning, related to the physics of the signal. Moreover, as the interactions between such structures are easily defined, the inference of higher level signal configurations can be made intuitive. The proposed algorithm improves the performance of existing video structures trackers, while reducing the computational complexity. The algorithm is demonstrated on audio-visual source localization.

In this work we present a method to jointly separate active audio and visual structures on a given mixture. Blind audiovisual source separation is achieved exploiting the coherence between a video signal and a one-microphone audio track. The efficient representation of audio and video sequences allows to build relationships between correlated structures on both modalities. Video structures exhibiting strong correlations with the audio signal and that are spatially close are grouped using a robust clustering algorithm that can count and localize audiovisual sources. Using such information and exploiting audio-video correlation, audio sources are also localized and separated. To the best of our knowledge this is the first blind audiovisual source separation algorithm conceived to deal with a video sequence and the correspondingmonoaudio signal.

Real-world phenomena involve complex interactions between multiple signal modalities. As a consequence, humans are used to integrate at each instant perceptions from all their senses in order to enrich their understanding of the surrounding world. This paradigm can be also extremely useful in many signal processing and computer vision problems involving mutually related signals. The simultaneous processing of multimodal data can, in fact, reveal information that is otherwise hidden when considering the signals independently. However, in natural multimodal signals, the statistical dependencies between modalities are in general not obvious. Learning fundamental multimodal patterns could offer deep insight into the structure of such signals. In this paper, we present a novel model of multimodal signals based on their sparse decomposition over a dictionary of multimodal structures. An algorithm for iteratively learning multimodal generating functions that can be shifted at all positions in the signal is proposed, as well. The learning is defined in such a way that it can be accomplished by iteratively solving a generalized eigenvector problem, which makes the algorithm fast, flexible, and free of user-defined parameters. The proposed algorithm is applied to audiovisual sequences and it is able to discover underlying structures in the data. The detection of such audio-video patterns in audiovisual clips allows to effectively localize the sound source on the video in presence of substantial acoustic and visual distractors, outperforming state-of-the-art audiovisual localization algorithms.

In this work we explore the potentialities of a framework for the representation of audio-visual signals using decompositions on overcomplete dictionaries. Redundant decompositions may describe audio-visual sequences in a concise fashion, preserving good representation properties thanks to the use of redundant, well designed, dictionaries. We expect that this helps us overcome two typical problems of multimodal fusion algorithms. On one hand, classical representation techniques, like pixel-based measures (for the video) or Fourier-like transforms (for the audio), take into account only marginally the physics of the problem. On the other hand, the input signals have large dimensionality. The results we obtain by making use of sparse decompositions of audio-visual signals over redundant codebooks are encouraging and show the potentialities of the proposed approach to multimodal signal representation.

The dictionary approach to signal and image processing has been massively investigated in the last two decades, proving very attractive for a wide range of applications. The effectiveness of dictionary-based methods, however, is strongly influenced by the choice of the set of basis functions. Moreover, the structure of the dictionary is of paramount importance regarding efficient implementation and practical applications such as image coding. In this work, an over-complete code for sparse representation of natural images has been learnt from a set a real-world scenes. The functions found have been organized into a hierarchical structure. We take advantage of this representation of the dictionary, adopting a structured greedy algorithm to build sparse approximations of images. Using this procedure, no a-priori constraint is imposed on the structure of the dictionary, allowing great flexibility in its design and lower computational complexity.

In the present paper we propose a new framework for the construction of meaningful dictionaries for sparse representation of signals. The dictionary approach to coding and compression proves very attractive since decomposing a signal over a redundant set of basis functions allows a parsimonious representation of information. This interest is witnessed by numerous research efforts that have been done in the last years to develop efficient algorithm for the decomposition of signals over redundant sets of functions. However, the effectiveness of such methods strongly depends on the dictionary and on its structure. In this work, we develop a method to learn overcomplete sets of functions from real-world signals. This technique allows the design of dictionaries that can be adapted to a specific class of signals. The found functions are stored in a tree structure. This data structure is used by a Tree-Based Pursuit algorithm to generate sparse approximations of natural signals. Finally, the proposed method is considered in the context of image compression. Results show that the learning Tree-Based approach outperforms state-of-the-art coding technique.