Neural radiance fields are an emerging 3D scene representation and recently even been extended to learn features for scene understanding by distilling open-vocabulary features from vision-language models. However, current method primarily focus on object-centric representations, supporting object segmentation or detection, while understanding semantic relationships between objects remains largely unexplored. To address this gap, we propose RelationField, the first method to extract inter-object relationships directly from neural radiance fields. RelationField represents relationships between objects as pairs of rays within a neural radiance field, effectively extending its formulation to include implicit relationship queries. To teach RelationField complex, open-vocabulary relationships, relationship knowledge is distilled from multi-modal LLMs. To evaluate RelationField, we solve open-vocabulary 3D scene graph generation tasks and relationship-guided instance segmentation, achieving state-of-the-art performance in both tasks. See the project website at relationfield.github.io.

3D scene graphs are an emerging 3D scene representation, that models both the objects present in the scene as well as their relationships. However, learning 3D scene graphs is a challenging task because it requires not only object labels but also relationship annotations, which are very scarce in datasets. While it is widely accepted that pre-training is an effective approach to improve model performance in low data regimes, in this paper, we find that existing pre-training methods are ill-suited for 3D scene graphs. To solve this issue, we present the first language-based pre-training approach for 3D scene graphs, whereby we exploit the strong relationship between scene graphs and language. To this end, we leverage the language encoder of CLIP, a popular vision-language model, to distill its knowledge into our graph-based network. We formulate a contrastive pre-training, which aligns text embeddings of relationships (subject-predicate-object triplets) and predicted 3D graph features. Our method achieves state-of-the-art results on the main semantic 3D scene graph benchmark by showing improved effectiveness over pre-training baselines and outperforming all the existing fully supervised scene graph prediction methods by a significant margin. Furthermore, since our scene graph features are language-aligned, it allows us to query the language space of the features in a zero-shot manner. In this paper, we show an example of utilizing this property of the features to predict the room type of a scene without further training.

Current approaches for 3D scene graph prediction rely on labeled datasets to train models for a fixed set of known object classes and relationship categories. We present Open3DSG, an alternative approach to learn 3D scene graph prediction in an open world without requiring labeled scene graph data. We co-embed the features from a 3D scene graph prediction backbone with the feature space of pow-erful open world 2D vision language foundation models. This enables us to predict 3D scene graphs from 3D point clouds in a zero-shot manner by querying object classes from an open vocabulary and predicting the inter-object relationships from a grounded LLM with scene graph features and queried object classes as context. Open3DSG is the first 3D point cloud method to predict not only explicit open-vocabulary object classes, but also open-set relation-ships that are not limited to a predefined label set, making it possible to express rare as well as specific objects and relationships in the predicted 3D scene graph. Our exper-iments show that Open3DSG is effective at predicting arbitrary object classes as well as their complex inter-object relationships describing spatial, supportive, semantic and comparative relationships.

Robust and accurate point cloud registration is an essential part of many robotic tasks such as SLAM or object pose retrieval. In this paper, we address the problem of global 3D point cloud registration, i.e., the task of estimating the 3D rigid body transform between a source and a target point cloud without any initial guess. Typically, the problem is solved by extracting and matching features to find a data association and then computing a transform that minimizes the squared distance between points. Our approach combines the normal distributions transform and oriented point pair framework and introduces the NDT distance histogram to quickly generate and test candidate transforms. Our method further exploits semantic information if available for greater speed. We implement our algorithm in C++ and compare it to other state-of-the-art approaches on a diverse set of environments. Our evaluation shows that our method outperforms the other approaches, especially concerning run-time and compute efficiency.

Visual Reinforcement Learning (RL) methods often require extensive amounts of data. As opposed to model-free RL, model-based RL (MBRL) offers a potential solution with efficient data utilization through planning. Additionally, RL lacks generalization capabilities for real-world tasks. Prior work has shown that incorporating pre-trained visual representations (PVRs) enhances sample efficiency and generalization. While PVRs have been extensively studied in the context of model-free RL, their potential in MBRL remains largely unexplored. In this paper, we benchmark a set of PVRs on challenging control tasks in a model-based RL setting. We investigate the data efficiency, generalization capabilities, and the impact of different properties of PVRs on the performance of model-based agents. Our results, perhaps surprisingly, reveal that for MBRL current PVRs are not more sample efficient than learning representations from scratch, and that they do not generalize better to out-of-distribution (OOD) settings. To explain this, we analyze the quality of the trained dynamics model. Furthermore, we show that data diversity and network architecture are the most important contributors to OOD generalization performance.

In the field of 3D scene understanding, 3D scene graphs have emerged as a new scene representation that combines geometric and semantic information about objects and their relationships. However, learning semantic 3D scene graphs in a fully supervised manner is inherently difficult as it requires not only object-level annotations but also relationship labels. While pre-training approaches have helped to boost the performance of many methods in various fields, pre-training for 3D scene graph prediction has received little attention. Furthermore, we find in this paper that classical contrastive point cloud-based pre-training approaches are ineffective for 3D scene graph learning. To this end, we present SGRec3D, a novel self-supervised pre-training method for 3D scene graph prediction. We propose to reconstruct the 3D input scene from a graph bottleneck as a pretext task. Pre-training SGRec3D does not require object relationship labels, making it possible to exploit large-scale 3D scene understanding datasets, which were off-limits for 3D scene graph learning before. Our experiments demonstrate that in contrast to recent point cloud-based pre-training approaches, our proposed pre-training improves the 3D scene graph prediction considerably, which results in SOTA performance, outperforming other 3D scene graph models by +10% on object prediction and +4% on relationship prediction. Additionally, we show that only using a small subset of 10% labeled data during fine-tuning is sufficient to outperform the same model without pre-training.

We investigate the task of object goal navigation in unknown environments where a target object is given as a semantic label (e.g. find a couch). This task is challenging as it requires the robot to consider the semantic context in diverse settings (e.g. TVs are often nearby couches). Most of the prior work tackles this problem under the assumption of a discrete action policy whereas we present an approach with continuous control which brings it closer to real world applications. In this paper, we use information-theoretic model predictive control on dense cost maps to bring object goal navigation closer to real robots with kinodynamic constraints. We propose a deep neural network framework to learn cost maps that encode semantic context and guide the robot towards the target object. We also present a novel way of fusing mid-level visual representations in our architecture to provide additional semantic cues for cost map prediction. The experiments show that our method leads to more efficient and accurate goal navigation with higher quality paths than the reported baselines. The results also indicate the importance of mid-level representations for navigation by improving the success rate by 8 percentage points.

Advances in sensing and learning algorithms have led to increasingly mature solutions for human detection by robots, particularly in selected use-cases such as pedestrian detection for self-driving cars or close-range person detection in consumer settings. Despite this progress, the simple question which sensor-algorithm combination is best suited for a person detection task at handƒ remains hard to answer. In this paper, we tackle this issue by conducting a systematic cross-modal analysis of sensor-algorithm combinations typically used in robotics. We compare the performance of state-of-the-art person detectors for 2D range data, 3D lidar, and RGB-D data as well as selected combinations thereof in a challenging industrial use-case.We further address the related problems of data scarcity in the industrial target domain, and that recent research on human detection in 3D point clouds has mostly focused on autonomous driving scenarios. To leverage these methodological advances for robotics applications, we utilize a simple, yet effective multi-sensor transfer learning strategy by extending a strong image-based RGB-D detector to provide cross-modal supervision for lidar detectors in the form of weak 3D bounding box labels.Our results show a large variance among the different approaches in terms of detection performance, generalization, frame rates and computational requirements. As our use-case contains difficulties representative for a wide range of service robot applications, we believe that these results point to relevant open challenges for further research and provide valuable support to practitioners for the design of their robot system.

We introduce a scalable approach for object pose estimation trained on simulated RGB views of multiple 3D models together. We learn an encoding of object views that does not only describe an implicit orientation of all objects seen during training, but can also relate views of untrained objects. Our single-encoder-multi-decoder network is trained using a technique we denote ”multi-path learning”: While the encoder is shared by all objects, each decoder only reconstructs views of a single object. Consequently, views of different instances do not have to be separated in the latent space and can share common features. The resulting encoder generalizes well from synthetic to real data and across various instances, categories, model types and datasets. We systematically investigate the learned encodings, their generalization, and iterative refinement strategies on the ModelNet40 and T-LESS dataset. Despite training jointly on multiple objects, our 6D Object Detection pipeline achieves state-of-the-art results on T-LESS at much lower runtimes than competing approaches.

While 2D object detection has made significant progress, robustly localizing objects in 3D space under presence of occlusion is still an unresolved issue. Our focus in this work is on real-time detection of human 3D centroids in RGB-D data. We propose an image-based detection approach which extends the YOLO v3 architecture with a 3D centroid loss and mid-level feature fusion to exploit complementary information from both modalities. We employ a transfer learning scheme which can benefit from existing large-scale 2D object detection datasets, while at the same time learning end-to-end 3D localization from our highly randomized, diverse synthetic RGB-D dataset with precise 3D groundtruth. We further propose a geometrically more accurate depth-aware crop augmentation for training on RGB-D data, which helps to improve 3D localization accuracy. In experiments on our challenging intralogistics dataset, we achieve state-of-the-art performance even when learning 3D localization just from synthetic data.

Superquadric surfaces play an important role in many different research fields due to their ability to model a variety of shapes with a small number of parameters. One of their core applications is the estimation of object shape characteristics from a set of discrete samples from the surface of an object. However, the corresponding optimization problem is prone to numerical instabilities in some regions of the parameter space. To mitigate this problem, lower bound constraints to the shape parameters are applied during the optimization thus limiting the range of shapes which can be accurately represented by superquadrics. Therefore, the exact modeling of very common shapes like cuboids and cylinders is error-prone in practice. In this article we investigate this problem and provide a numerically stable formulation for the evaluation of the superquadric surface function and for its gradient. This new formulation enables numerically stable fitting of superquadrics in the previously constrained region, i.e., in its full range including cuboids and cylinders. In addition, the new formulation also leads to faster convergence speed. The theoretical contributions are substantiated by experiments on synthetic as well as real data.

An extension based on attention and anticipation of a robot vision pipeline for object recognition in RGBD images from low-cost sensors like MS Kinect or ASUS Xtion is presented. This work originated in research on an industrial application scenario, namely shipping-container unloading, but it is applicable to advanced manipulation tasks in unstructured environments in general where the perception must be very robust while being as fast as possible. For these scenarios, we build on our previous work that proved to be competitive in cluttered scenes in table-top scenarios and which forms the backbone of our RGBD object recognition. It is further enhanced by two main contributions. First, a simple but very effective form of anticipation as top-down expectations of the evolution of the scene due to the actions of the robot is used to speed up the processing. Second, attention is used as a mechanism for further speed-up by focusing processing only on certain regions of interest of the scene based also on an anticipation mechanism. The method is analyzed in experiments using real-world data from an industrial demonstration set-up.

Given that 3D scan matching is such a central part of the perception pipeline for robots, thorough and large-scale investigations of scan matching performance are still surprisingly few. A crucial part of the scientific method is to perform experiments that can be replicated by other researchers in order to compare different results. In light of this fact, this paper presents a thorough comparison of 3D scan registration algorithms using a recently published benchmark protocol which makes use of a publicly available challenging data set that covers a wide range of environments. In particular, we evaluate two types of recent 3D registration algorithms - one local and one global. Both approaches take local surface structure into account, rather than matching individual points. After well over 100 000 individual tests, we conclude that algorithms using the normal distributions transform (NDT) provides accurate results compared to a modern implementation of the iterative closest point (ICP) method, when faced with scan data that has little overlap and weak geometric structure. We also demonstrate that the minimally uncertain maximum consensus (MUMC) algorithm provides accurate results in structured environments without needing an initial guess, and that it provides useful measures to detect whether it has succeeded or not. We also propose two amendments to the experimental protocol, in order to provide more valuable results in future implementations.

In this work we tackle the problem of planning grasps for an underactuated gripper which enable it to retrieve target objects from a cluttered environment. Furthermore, we investigate how additional manipulation capabilities of the gripping device, provided by active surfaces on the inside of the fingers, can lead to performance improvement in the grasp execution process. To this end, we employ a simple strategy, in which the target object is `pulled-in' towards the palm during grasping which results in firm enveloping grasps. We show the effectiveness of the suggested methods by means of experiments conducted in a real-world scenario.

This paper presents an approach to acquire textured 3D models of objects without the need for sophisticated hardware infrastructures. The approach is inexpensive, using a low-cost Microsoft Kinect RGB-D sensor and Augmented Reality (AR) markers printed on paper sheets. The AR-markers can be freely placed in the scene, allowing the modeling of objects of various sizes, and the sensor can be moved by the hand of an untrained person. To generate usable models with this very inexpensive and simple setup, the sequence of RGB-D scans is embedded in a graph-based optimizer for automatic post-refinement. The main novelty of this contribution is the development of an uncertainty model for an AR-marker. The AR-marker uncertainty models are used as constraints in an optimization problem to better estimate the object pose. The models are in the end further fine-tuned by a standard point-based registration algorithm. The results section presents realistic models of various objects generated using this system, e.g., parcels, sport balls, human dolls etc. Additionally, a quantitative analysis is presented using objects of known dimensions.

In this paper, we give an overview of the Jacobs Robotics entry to the ICRA'11 Solutions in Perception Challenge. We present our multi-pronged strategy for object recognition and localization based on the integrated geometric and visual information available from the Kinect sensor. Firstly, the range image is over-segmented using an edge-detection algorithm and regions of interest are extracted based on a simple shape-analysis per segment. Then, these selected regions of the scene are matched with known objects using visual features and their distribution in 3D space. Finally, generated hypotheses about the positions of the objects are tested by back-projecting learned 3D models to the scene using estimated transformations and sensor model. Our method won the second place among eight competing algorithms, only marginally losing to the winner.

Fitting optimum quadrics on segmented noisy range-image is a challenging task. The algebraic least-squares fit usually considered in the literature is biased, although it is fast to compute. We extend our previous work in planar patch extraction using a detailed sensor noise model to the extraction of quadrics. First, the fast least-squares method is modified to detect degeneracy automatically and it is used to aid segmentation. The final segmented quadric patch is then refined using a numerical maximum likelihood approach which employs the sensor range error model. Experimental results for artificial data and two different 3D sensors are provided to show the feasibility of our approach.

The recently introduced Minimum Uncertainty Maximum Consensus (MUMC) algorithm for 3D scene registration using planar-patches is tested in a large outdoor urban setting without any prior motion estimate whatsoever. With the aid of a new overlap metric based on unmatched patches, the algorithm is shown to work successfully in most cases. The absolute accuracy of its computed result is corroborated for the first time by ground-truth obtained using reflective markers. There were a couple of unsuccessful scan-pairs. These are analyzed for the reason of failure by formulating two kinds of overlap metrics: one based on the actual overlapping surface-area and another based on the extent of agreement of range-image pixels. We conclude that neither metric in isolation is able to predict all failures, but that both taken together are able to predict the difficulty level of a scan-pair vis-à-vis registration by MUMC.

Surface-patches based 3D mapping in a real world underwater scenario is presented. It is based on a 6 degrees of freedom registration of sonar data. Planar surfaces are fitted into the sonar data and the subsequent registration method maximizes the overall geometric consistency within a search-space to determine correspondences between the planes. This approach has previously only been used on high quality range data from sensors on land robots like laser range finders. It is shown here that the algorithm is also applicable to very noisy, coarse sonar data. The 3D map presented is of a large underwater structure, namely the Lesumer Sperrwerk, a flood gate north of the city of Bremen, Germany. It is generated from 18 scans collected using a Tritech Eclipse sonar.

We present a robot-pose-registration algorithm, which is entirely based on large planar-surface patches extracted from point clouds sampled from a three-dimensional (3-D) sensor. This approach offers an alternative to the traditional point-to-point iterative-closest-point (ICP) algorithm, its point-to-plane variant, as well as newer grid-based algorithms, such as the 3-D normal distribution transform (NDT). The simpler case of known plane correspondences is tackled first by deriving expressions for least-squares pose estimation considering plane-parameter uncertainty computed during plane extraction. Closed-form expressions for covariances are also derived. To round-off the solution, we present a new algorithm, which is called minimally uncertain maximal consensus (MUMC), to determine the unknown plane correspondences by maximizing geometric consistency by minimizing the uncertainty volume in configuration space. Experimental results from three 3-D sensors, viz., Swiss-Ranger, University of South Florida Odetics Laser Detection and Ranging, and an actuated SICK S300, are given. The first two have low fields of view (FOV) and moderate ranges, while the third has a much bigger FOV and range. Experimental results show that this approach is not only more robust than point- or grid-based approaches in plane-rich environments, but it is also faster, requires significantly less memory, and offers a less-cluttered planar-patches-based visualization.

In this work, we utilize a recently studied more accurate range noise model for 3D sensors to derive from scratch the expressions for the optimum plane which best fits a point-cloud and for the combined covariance matrix of the plane's parameters. The parameters in question are the plane's normal and its distance to the origin. The range standard-deviation model used by us is a quadratic function of the true range and is a function of the incidence angle as well. We show that for this model, the maximum-likelihood plane is biased, whereas the least-squares plane is not. The plane-parameters' covariance matrix for the least-squares plane is shown to possess a number of desirable properties, e.g., the optimal solution forms its null-space and its components are functions of easily understood terms like the planar-patch's center and scatter. We verify our covariance expression with that obtained by the eigenvector perturbation method. We further compare our method to that of renormalization with respect to the theoretically best covariance matrix in simulation. The application of our approach to real-time range-image registration and plane fusion is shown by an example using a commercially available 3D range sensor. Results show that our method has good accuracy, is fast to compute, and is easy to interpret intuitively.

This article addresses fast 3D mapping by a mobile robot in a predominantly planar environment. It is based on a novel pose registration algorithm based entirely on matching features composed of plane-segments extracted from point-clouds sampled from a 3D sensor. The approach has advantages in terms of robustness, speed and storage as compared to the voxel based approaches. Unlike previous approaches, the uncertainty in plane parameters is utilized to compute the uncertainty in the pose computed by scan-registration. The algorithm is illustrated by creating a full 3D model of a multi-level robot testing arena.

A fast but nevertheless accurate approach for surface extraction from noisy 3D point clouds is presented. It consists of two parts, namely a plane fitting and a polygonalization step. Both exploit the sequential nature of 3D data acquisition on mobile robots in form of range images. For the plane fitting, this is used to revise the standard mathematical formulation to an incremental version, which allows a linear computation. For the polygonalization, the neighborhood relation in range images is exploited. Experiments are presented using a time-of-flight range camera in form of a Swissranger SR-3000. Results include lab scenes as well as data from two runs of the rescue robot league at the RoboCup German Open 2007 with 1,414, respectively 2,343 sensor snapshots. The 36ldr106, respectively 59ldr106points from the two point clouds are reduced to about 14ldr103, respectively 23ldr103planes with only about 0.2 sec of total computation time per snapshot while the robot moves along. Uncertainty analysis of the computed plane parameters is presented as well.