Stefano Rosa

Tommaso Galliena, Stefano Rosa, Tommaso Apicella et al.

Vision-Language Models (VLMs) often yield inconsistent descriptions of the same object across viewpoints, hindering the ability of embodied agents to construct consistent semantic representations over time. Previous methods resolved inconsistencies using offline multi-view aggregation or multi-stage pipelines that decouple exploration, data association, and caption learning, with limited capacity to reason over previously observed objects. In this paper, we introduce a unified, memory-augmented Vision-Language agent that simultaneously handles data association, object captioning, and exploration policy within a single autoregressive framework. The model processes the current RGB observation, a top-down explored map, and an object-level episodic memory serialized into object-level tokens, ensuring persistent object identity and semantic consistency across extended sequences. To train the model in a self-supervised manner, we collect a dataset in photorealistic 3D environments using a disagreement-based policy and a pseudo-captioning model that enforces consistency across multi-view caption histories. Extensive evaluation on a manually annotated object-level test set, demonstrate improvements of up to +11.86% in standard captioning scores and +7.39% in caption self-similarity over baseline models, while enabling scalable performance through a compact scene representation. Code, model weights, and data are available at https://github.com/hsp-iit/epos-vlm

Gianluca Scarpellini, Stefano Rosa, Pietro Morerio et al.

When an object detector is deployed in a novel setting it often experiences a drop in performance. This paper studies how an embodied agent can automatically fine-tune a pre-existing object detector while exploring and acquiring images in a new environment without relying on human intervention, i.e., a fully self-supervised approach. In our setting, an agent initially learns to explore the environment using a pre-trained off-the-shelf detector to locate objects and associate pseudo-labels. By assuming that pseudo-labels for the same object must be consistent across different views, we learn the exploration policy Look Around to mine hard samples, and we devise a novel mechanism called Disagreement Reconciliation for producing refined pseudo-labels from the consensus among observations. We implement a unified benchmark of the current state-of-the-art and compare our approach with pre-existing exploration policies and perception mechanisms. Our method is shown to outperform existing approaches, improving the object detector by 6.2% in a simulated scenario, a 3.59% advancement over other state-of-the-art methods, and by 9.97% in the real robotic test without relying on ground-truth. Code for the proposed approach and baselines are available at https://iit-pavis.github.io/Look_Around_And_Learn/.

Gianluca Scarpellini, Stefano Rosa, Pietro Morerio et al.

Object detectors often experience a drop in performance when new environmental conditions are insufficiently represented in the training data. This paper studies how to automatically fine-tune a pre-existing object detector while exploring and acquiring images in a new environment without relying on human intervention, i.e., in a self-supervised fashion. In our setting, an agent initially explores the environment using a pre-trained off-the-shelf detector to locate objects and associate pseudo-labels. By assuming that pseudo-labels for the same object must be consistent across different views, we devise a novel mechanism for producing refined predictions from the consensus among observations. Our approach improves the off-the-shelf object detector by 2.66% in terms of mAP and outperforms the current state of the art without relying on ground-truth annotations.

Francesco Taioli, Stefano Rosa, Alberto Castellini et al.

Vision-and-Language Navigation in Continuous Environments (VLN-CE) is one of the most intuitive yet challenging embodied AI tasks. Agents are tasked to navigate towards a target goal by executing a set of low-level actions, following a series of natural language instructions. All VLN-CE methods in the literature assume that language instructions are exact. However, in practice, instructions given by humans can contain errors when describing a spatial environment due to inaccurate memory or confusion. Current VLN-CE benchmarks do not address this scenario, making the state-of-the-art methods in VLN-CE fragile in the presence of erroneous instructions from human users. For the first time, we propose a novel benchmark dataset that introduces various types of instruction errors considering potential human causes. This benchmark provides valuable insight into the robustness of VLN systems in continuous environments. We observe a noticeable performance drop (up to -25%) in Success Rate when evaluating the state-of-the-art VLN-CE methods on our benchmark. Moreover, we formally define the task of Instruction Error Detection and Localization, and establish an evaluation protocol on top of our benchmark dataset. We also propose an effective method, based on a cross-modal transformer architecture, that achieves the best performance in error detection and localization, compared to baselines. Surprisingly, our proposed method has revealed errors in the validation set of the two commonly used datasets for VLN-CE, i.e., R2R-CE and RxR-CE, demonstrating the utility of our technique in other tasks. Code and dataset available at https://intelligolabs.github.io/R2RIE-CE

Tommaso Galliena, Tommaso Apicella, Stefano Rosa et al.

We present a self-supervised method to improve an agent's abilities in describing arbitrary objects while actively exploring a generic environment. This is a challenging problem, as current models struggle to obtain coherent image captions due to different camera viewpoints and clutter. We propose a three-phase framework to fine-tune existing captioning models that enhances caption accuracy and consistency across views via a consensus mechanism. First, an agent explores the environment, collecting noisy image-caption pairs. Then, a consistent pseudo-caption for each object instance is distilled via consensus using a large language model. Finally, these pseudo-captions are used to fine-tune an off-the-shelf captioning model, with the addition of contrastive learning. We analyse the performance of the combination of captioning models, exploration policies, pseudo-labeling methods, and fine-tuning strategies, on our manually labeled test set. Results show that a policy can be trained to mine samples with higher disagreement compared to classical baselines. Our pseudo-captioning method, in combination with all policies, has a higher semantic similarity compared to other existing methods, and fine-tuning improves caption accuracy and consistency by a significant margin. Code and test set annotations available at https://hsp-iit.github.io/embodied-captioning/

Francesco Taioli, Stefano Rosa, Alberto Castellini et al.

In the Vision-and-Language Navigation in Continuous Environments (VLN-CE) task, the human user guides an autonomous agent to reach a target goal via a series of low-level actions following a textual instruction in natural language. However, most existing methods do not address the likely case where users may make mistakes when providing such instruction (e.g. "turn left" instead of "turn right"). In this work, we address a novel task of Interactive VLN in Continuous Environments (IVLN-CE), which allows the agent to interact with the user during the VLN-CE navigation to verify any doubts regarding the instruction errors. We propose an Interactive Instruction Error Detector and Localizer (I2EDL) that triggers the user-agent interaction upon the detection of instruction errors during the navigation. We leverage a pre-trained module to detect instruction errors and pinpoint them in the instruction by cross-referencing the textual input and past observations. In such way, the agent is able to query the user for a timely correction, without demanding the user's cognitive load, as we locate the probable errors to a precise part of the instruction. We evaluate the proposed I2EDL on a dataset of instructions containing errors, and further devise a novel metric, the Success weighted by Interaction Number (SIN), to reflect both the navigation performance and the interaction effectiveness. We show how the proposed method can ask focused requests for corrections to the user, which in turn increases the navigation success, while minimizing the interactions.

Qingyong Hu, Bo Yang, Linhai Xie et al.

We study the problem of efficient semantic segmentation of large-scale 3D point clouds. By relying on expensive sampling techniques or computationally heavy pre/post-processing steps, most existing approaches are only able to be trained and operate over small-scale point clouds. In this paper, we introduce RandLA-Net, an efficient and lightweight neural architecture to directly infer per-point semantics for large-scale point clouds. The key to our approach is to use random point sampling instead of more complex point selection approaches. Although remarkably computation and memory efficient, random sampling can discard key features by chance. To overcome this, we introduce a novel local feature aggregation module to progressively increase the receptive field for each 3D point, thereby effectively preserving geometric details. Comparative experiments show that our RandLA-Net can process 1 million points in a single pass up to 200x faster than existing approaches. Moreover, extensive experiments on five large-scale point cloud datasets, including Semantic3D, SemanticKITTI, Toronto3D, NPM3D and S3DIS, demonstrate the state-of-the-art semantic segmentation performance of our RandLA-Net.

Changhao Chen, Stefano Rosa, Chris Xiaoxuan Lu et al.

Autonomous vehicles and mobile robotic systems are typically equipped with multiple sensors to provide redundancy. By integrating the observations from different sensors, these mobile agents are able to perceive the environment and estimate system states, e.g. locations and orientations. Although deep learning approaches for multimodal odometry estimation and localization have gained traction, they rarely focus on the issue of robust sensor fusion - a necessary consideration to deal with noisy or incomplete sensor observations in the real world. Moreover, current deep odometry models suffer from a lack of interpretability. To this extent, we propose SelectFusion, an end-to-end selective sensor fusion module which can be applied to useful pairs of sensor modalities such as monocular images and inertial measurements, depth images and LIDAR point clouds. Our model is a uniform framework that is not restricted to specific modality or task. During prediction, the network is able to assess the reliability of the latent features from different sensor modalities and estimate trajectory both at scale and global pose. In particular, we propose two fusion modules - a deterministic soft fusion and a stochastic hard fusion, and offer a comprehensive study of the new strategies compared to trivial direct fusion. We extensively evaluate all fusion strategies in both public datasets and on progressively degraded datasets that present synthetic occlusions, noisy and missing data and time misalignment between sensors, and we investigate the effectiveness of the different fusion strategies in attending the most reliable features, which in itself, provides insights into the operation of the various models.

Qingyong Hu, Bo Yang, Linhai Xie et al.

We study the problem of efficient semantic segmentation for large-scale 3D point clouds. By relying on expensive sampling techniques or computationally heavy pre/post-processing steps, most existing approaches are only able to be trained and operate over small-scale point clouds. In this paper, we introduce RandLA-Net, an efficient and lightweight neural architecture to directly infer per-point semantics for large-scale point clouds. The key to our approach is to use random point sampling instead of more complex point selection approaches. Although remarkably computation and memory efficient, random sampling can discard key features by chance. To overcome this, we introduce a novel local feature aggregation module to progressively increase the receptive field for each 3D point, thereby effectively preserving geometric details. Extensive experiments show that our RandLA-Net can process 1 million points in a single pass with up to 200X faster than existing approaches. Moreover, our RandLA-Net clearly surpasses state-of-the-art approaches for semantic segmentation on two large-scale benchmarks Semantic3D and SemanticKITTI.

Muhamad Risqi U. Saputra, Pedro P. B. de Gusmao, Chris Xiaoxuan Lu et al.

Visual odometry shows excellent performance in a wide range of environments. However, in visually-denied scenarios (e.g. heavy smoke or darkness), pose estimates degrade or even fail. Thermal cameras are commonly used for perception and inspection when the environment has low visibility. However, their use in odometry estimation is hampered by the lack of robust visual features. In part, this is as a result of the sensor measuring the ambient temperature profile rather than scene appearance and geometry. To overcome this issue, we propose a Deep Neural Network model for thermal-inertial odometry (DeepTIO) by incorporating a visual hallucination network to provide the thermal network with complementary information. The hallucination network is taught to predict fake visual features from thermal images by using Huber loss. We also employ selective fusion to attentively fuse the features from three different modalities, i.e thermal, hallucination, and inertial features. Extensive experiments are performed in hand-held and mobile robot data in benign and smoke-filled environments, showing the efficacy of the proposed model.

Changhao Chen, Stefano Rosa, Yishu Miao et al.

Deep learning approaches for Visual-Inertial Odometry (VIO) have proven successful, but they rarely focus on incorporating robust fusion strategies for dealing with imperfect input sensory data. We propose a novel end-to-end selective sensor fusion framework for monocular VIO, which fuses monocular images and inertial measurements in order to estimate the trajectory whilst improving robustness to real-life issues, such as missing and corrupted data or bad sensor synchronization. In particular, we propose two fusion modalities based on different masking strategies: deterministic soft fusion and stochastic hard fusion, and we compare with previously proposed direct fusion baselines. During testing, the network is able to selectively process the features of the available sensor modalities and produce a trajectory at scale. We present a thorough investigation on the performances on three public autonomous driving, Micro Aerial Vehicle (MAV) and hand-held VIO datasets. The results demonstrate the effectiveness of the fusion strategies, which offer better performances compared to direct fusion, particularly in presence of corrupted data. In addition, we study the interpretability of the fusion networks by visualising the masking layers in different scenarios and with varying data corruption, revealing interesting correlations between the fusion networks and imperfect sensory input data.

Linhai Xie, Sen Wang, Stefano Rosa et al.

Deep Reinforcement Learning (DRL) has been applied successfully to many robotic applications. However, the large number of trials needed for training is a key issue. Most of existing techniques developed to improve training efficiency (e.g. imitation) target on general tasks rather than being tailored for robot applications, which have their specific context to benefit from. We propose a novel framework, Assisted Reinforcement Learning, where a classical controller (e.g. a PID controller) is used as an alternative, switchable policy to speed up training of DRL for local planning and navigation problems. The core idea is that the simple control law allows the robot to rapidly learn sensible primitives, like driving in a straight line, instead of random exploration. As the actor network becomes more advanced, it can then take over to perform more complex actions, like obstacle avoidance. Eventually, the simple controller can be discarded entirely. We show that not only does this technique train faster, it also is less sensitive to the structure of the DRL network and consistently outperforms a standard Deep Deterministic Policy Gradient network. We demonstrate the results in both simulation and real-world experiments.

Zhihua Wang, Stefano Rosa, Yishu Miao et al.

Humans are able to make rich predictions about the future dynamics of physical objects from a glance. On the other hand, most existing computer vision approaches require strong assumptions about the underlying system, ad-hoc modeling, or annotated datasets, to carry out even simple predictions. To tackle this gap, we propose a new perspective on the problem of learning intuitive physics that is inspired by the spatial memory representation of objects and spaces in human brains, in particular the co-existence of egocentric and allocentric spatial representations. We present a generic framework that learns a layered representation of the physical world, using a cascade of invertible modules. In this framework, real images are first converted to a synthetic domain representation that reduces complexity arising from lighting and texture. Then, an allocentric viewpoint transformer removes viewpoint complexity by projecting images to a canonical view. Finally, a novel Recurrent Latent Variation Network (RLVN) architecture learns the dynamics of the objects interacting with the environment and predicts future motion, leveraging the availability of unlimited synthetic simulations. Predicted frames are then projected back to the original camera view and translated back to the real world domain. Experimental results show the ability of the framework to consistently and accurately predict several frames in the future and the ability to adapt to real images.

Zhihua Wang, Stefano Rosa, Bo Yang et al.

The ability to interact and understand the environment is a fundamental prerequisite for a wide range of applications from robotics to augmented reality. In particular, predicting how deformable objects will react to applied forces in real time is a significant challenge. This is further confounded by the fact that shape information about encountered objects in the real world is often impaired by occlusions, noise and missing regions e.g. a robot manipulating an object will only be able to observe a partial view of the entire solid. In this work we present a framework, 3D-PhysNet, which is able to predict how a three-dimensional solid will deform under an applied force using intuitive physics modelling. In particular, we propose a new method to encode the physical properties of the material and the applied force, enabling generalisation over materials. The key is to combine deep variational autoencoders with adversarial training, conditioned on the applied force and the material properties. We further propose a cascaded architecture that takes a single 2.5D depth view of the object and predicts its deformation. Training data is provided by a physics simulator. The network is fast enough to be used in real-time applications from partial views. Experimental results show the viability and the generalisation properties of the proposed architecture.

Zhihua Wang, Stefano Rosa, Linhai Xie et al.

Modelling the physical properties of everyday objects is a fundamental prerequisite for autonomous robots. We present a novel generative adversarial network (Defo-Net), able to predict body deformations under external forces from a single RGB-D image. The network is based on an invertible conditional Generative Adversarial Network (IcGAN) and is trained on a collection of different objects of interest generated by a physical finite element model simulator. Defo-Net inherits the generalisation properties of GANs. This means that the network is able to reconstruct the whole 3-D appearance of the object given a single depth view of the object and to generalise to unseen object configurations. Contrary to traditional finite element methods, our approach is fast enough to be used in real-time applications. We apply the network to the problem of safe and fast navigation of mobile robots carrying payloads over different obstacles and floor materials. Experimental results in real scenarios show how a robot equipped with an RGB-D camera can use the network to predict terrain deformations under different payload configurations and use this to avoid unsafe areas.

Bo Yang, Stefano Rosa, Andrew Markham et al.

In this paper, we propose a novel approach, 3D-RecGAN++, which reconstructs the complete 3D structure of a given object from a single arbitrary depth view using generative adversarial networks. Unlike existing work which typically requires multiple views of the same object or class labels to recover the full 3D geometry, the proposed 3D-RecGAN++ only takes the voxel grid representation of a depth view of the object as input, and is able to generate the complete 3D occupancy grid with a high resolution of 256^3 by recovering the occluded/missing regions. The key idea is to combine the generative capabilities of autoencoders and the conditional Generative Adversarial Networks (GAN) framework, to infer accurate and fine-grained 3D structures of objects in high-dimensional voxel space. Extensive experiments on large synthetic datasets and real-world Kinect datasets show that the proposed 3D-RecGAN++ significantly outperforms the state of the art in single view 3D object reconstruction, and is able to reconstruct unseen types of objects.

Giorgio Toscana, Stefano Rosa

Object segmentation is an important capability for robotic systems, in particular for grasping. We present a graph- based approach for the segmentation of simple objects from RGB-D images. We are interested in segmenting objects with large variety in appearance, from lack of texture to strong textures, for the task of robotic grasping. The algorithm does not rely on image features or machine learning. We propose a modified Canny edge detector for extracting robust edges by using depth information and two simple cost functions for combining color and depth cues. The cost functions are used to build an undirected graph, which is partitioned using the concept of internal and external differences between graph regions. The partitioning is fast with O(NlogN) complexity. We also discuss ways to deal with missing depth information. We test the approach on different publicly available RGB-D object datasets, such as the Rutgers APC RGB-D dataset and the RGB-D Object Dataset, and compare the results with other existing methods.