ICRA 2023 Papers

Manipulating cables is challenging for robots because of the infinite degrees of freedom of the cables and frequent occlusion by the gripper and the environment. These challenges are further complicated by the dexterous nature of the operations required for cable routing and assembly, such as weaving and inserting, hampering common solutions with vision-only sensing. In this paper, we propose to integrate tactile-guided low-level motion control with high-level vision- based task parsing for a challenging task: cable routing and assembly on a reconfigurable task board. Specifically, we build a library of tactile-guided motion primitives using a fingertip GelSight sensor, where each primitive reliably accomplishes an operation such as cable following and weaving. The overall task is inferred via visual perception given a goal configuration image, and then used to generate the primitive sequence. Experiments demonstrate the effectiveness of individual tactile- guided primitives and the integrated end-to-end solution, sig- nificantly outperforming the method without tactile sensing. Our reconfigurable task setup and proposed baselines provide a benchmark for future research in cable manipulation.

We address the important problem of generalizing robotic rearrangement to clutter without any explicit object models. We first generate over 650K cluttered scenes-orders of magnitude more than prior work-in diverse everyday environments, such as cabinets and shelves. We render synthetic partial point clouds from this data and use it to train our CabiNet model architecture. CabiNet is a collision model that accepts object and scene point clouds, captured from a single-view depth observation, and predicts collisions for SE(3) object poses in the scene. Our representation has a fast inference speed of 7μs/query with nearly 20% higher performance than baseline approaches in challenging environments. We use this collision model in conjunction with a Model Predictive Path Integral (MPPI) planner to generate collision-free trajectories for picking and placing in clutter. CabiNet also predicts waypoints, computed from the scene's signed distance field (SDF), that allows the robot to navigate tight spaces during rearrangement. This improves rearrangement performance by nearly 35% compared to baselines. We systematically evaluate our approach, procedurally generate simulated experiments, and demonstrate that our approach directly transfers to the real world, despite training exclusively in simulation. Supplementary material and videos of robot experiments in completely unknown scenes are available at: cabinet-object-rearrangement.github.io.

The current approach to exploring and monitoring complex underwater ecosystems, such as coral reefs, is to conduct surveys using diver-held or static cameras, or deploying sensor buoys. These approaches often fail to capture the full variation and complexity of interactions between different reef organisms and their habitat. The CUREE platform presented in this paper provides a unique set of capabilities in the form of robot behaviors and perception algorithms to enable scientists to explore different aspects of an ecosystem. Examples of these capabilities include low-altitude visual surveys, soundscape surveys, habitat characterization, and animal following. We demonstrate these capabilities by describing two field deployments on coral reefs in the US Virgin Islands. In the first deployment, we show that CUREE can identify the preferred habitat type of snapping shrimp in a reef through a combination of a visual survey, habitat characterization, and a soundscape survey. In the second deployment, we demonstrate CUREE's ability to follow arbitrary animals by separately following a barracuda and stingray for several minutes each in midwater and benthic environments, respectively.

The Perspective-n-Point (PnP) problem has been widely studied in both computer vision and photogrammetry societies. With the development of feature extraction techniques, a large number of feature points might be available in a single shot. It is promising to devise a consistent estimator, i.e., the estimate can converge to the true camera pose as the number of points increases. To this end, we propose a consistent PnP solver, named CPnP, with bias elimination. Specifically, linear equations are constructed from the original projection model via measurement model modification and variable elimination, based on which a closed-form least-squares solution is obtained. We then analyze and subtract the asymptotic bias of this solution, resulting in a consistent estimate. Additionally, Gauss-Newton (GN) iterations are executed to refine the consistent solution. Our proposed estimator is efficient in terms of computations—it has O(n)O(n) time complexity. Simulations and real dataset tests show that our proposed estimator is superior to some well-known ones for images with dense visual features, in terms of estimation precision and computing time.

A fast and accurate panoptic segmentation system for LiDAR point clouds is crucial for autonomous driving vehicles to understand the surrounding objects and scenes. Existing approaches usually rely on proposals or clustering to segment foreground instances. As a result, they struggle to achieve real-time performance. In this paper, we propose a novel real-time end-to-end panoptic segmentation network for LiDAR point clouds, called CPSeg. In particular, CPSeg comprises a shared encoder, a dual-decoder, and a cluster-free instance segmentation head, which is able to dynamically pillarize foreground points according to the learned embedding. Then, it acquires instance labels by finding connected pillars with a pairwise embedding comparison. Thus, the conventional proposal-based or clustering-based instance segmentation is transformed into a binary segmentation problem on the pairwise embedding comparison matrix. To help the network regress instance embedding, a fast and deterministic depth completion algorithm is proposed to calculate the surface normal of each point cloud in real-time. The proposed method is benchmarked on two large-scale autonomous driving datasets: SemanticKITTI and nuScenes. Notably, extensive experimental results show that CPSeg achieves state-of-the-art results among real-time approaches on both datasets.

Collaborative SLAM is at the core of perception in multi-robot systems as it enables the co-localization of the team of robots in a common reference frame, which is of vital importance for any coordination amongst them. The paradigm of a centralized architecture is well established, with the robots (i.e. agents) running Visual-Inertial Odometry (VIO) onboard while communicating relevant data, such as e.g. Keyframes (KFs), to a central back-end (i.e. server), which then merges and optimizes the joint maps of the agents. While these frameworks have proven to be successful, their capability and performance are highly dependent on the choice of the VIO front-end, thus limiting their flexibility. In this work, we present COVINSG, a generalized back-end building upon the COVINS [1] framework, enabling the compatibility of the server-back-end with any arbitrary VIO front-end, including, for example, off-the-shelf cameras with odometry capabilities, such as the Realsense T265. The COVINS-G back-end deploys a multi-camera relative pose estimation algorithm for computing the loop-closure constraints allowing the system to work purely on 2D image data. In the experimental evaluation, we show on-par accuracy with state-of-the-art multi-session and collaborative SLAM systems, while demonstrating the flexibility and generality of our approach by employing different front-ends onboard collaborating agents within the same mission. The COVINS-G codebase along with a generalized front-end wrapper to allow any existing VIO front-end to be readily used in combination with the proposed collaborative back-end is open-sourced. Video- https://youtu.be/FoJfXCfaYDw

Transfer learning is a proven technique in 2D computer vision to leverage the large amount of data available and achieve high performance with datasets limited in size due to the cost of acquisition or annotation. In 3D, annotation is known to be a costly task; nevertheless, pre-training methods have only recently been investigated. Due to this cost, unsupervised pretraining has been heavily favored. In this work, we tackle the case of real-time 3D semantic segmentation of sparse autonomous driving LiDAR scans. Such datasets have been increasingly released, but each has a unique label set. We propose here an intermediate-level label set called coarse labels, which can easily be used on any existing and future autonomous driving datasets, thus allowing all the data available to be leveraged at once without any additional manual labeling. This way, we have access to a larger dataset, alongside a simple task of semantic segmentation. With it, we introduce a new pretraining task: coarse label pre-training, also called COLA. We thoroughly analyze the impact of COLA on various datasets and architectures and show that it yields a noticeable performance improvement, especially when only a small dataset is available for the finetuning task.

This paper proposes CNN-based visual servoing for simultaneous positioning and flattening of a soft fabric part placed on a table by a dual manipulator system. We propose a network for multimodal data processing of grayscale images captured by a camera and force/torque applied to force sensors. The training dataset is collected by moving the real manipulators, which enables the network to map the captured images and force/torque to the manipulator's motion in Cartesian space. We apply structured lighting to emphasize the features of the surface of the fabric part since the surface shape of the non-textured fabric part is difficult to recognize by a single grayscale image. Through experiments, we show that the fabric part with unseen wrinkles can be positioned and flattened by the proposed visual servoing scheme.

In this paper, we propose a novel representation for grasping using contacts between multi-finger robotic hands and objects to be manipulated. This representation significantly reduces the prediction dimensions and accelerates the learning process. We present an effective end-to-end network, CMG-Net, for grasping unknown objects in a cluttered environment by efficiently predicting multi-finger grasp poses and hand configurations from a single-shot point cloud. Moreover, we create a synthetic grasp dataset that consists of five thousand cluttered scenes, 80 object categories, and 20 million annotations. We perform a comprehensive empirical study and demonstrate the effectiveness of our grasping representation and CMG-Net. Our work significantly outperforms the state-of-the-art for three-finger robotic hands. We also demonstrate that the model trained using synthetic data perform very well for real robots.

Rescue missions in mountain environments are hardly achievable by standard legged robots—because of the high slopes—or by flying robots—because of limited payload capacity. We present a concept for a rope-aided climbing robot which can negotiate up-to-vertical slopes and carry heavy payloads. The robot is attached to the mountain through a rope, and it is equipped with a leg to push against the mountain and initiate jumping maneuvers. Between jumps, a hoist is used to wind/unwind the rope to move vertically and affect the lateral motion. This simple (yet effective) two-fold actuation allows the system to achieve high safety and energy efficiency. Indeed, the rope prevents the robot from falling while compensating for most of its weight, drastically reducing the effort required by the leg actuator. We also present an optimal control strategy to generate point-to-point trajectories overcoming an obstacle. We achieve fast computation time (<1 s) thanks to the use of a custom simplified robot model. We validated the generated optimal movements in Gazebo simulations with a complete robot model with a < 5% error on a 16\ m16\ m long jump, showing the effectiveness of the proposed approach, and confirming the interest of our concept. Finally, we performed a reachability analysis showing that the region of achievable targets is strongly affected by the friction properties of the foot-wall contact.

The ongoing electrification in all domains relies on strong increase in raw material extraction. Autonomous dump trucks are key to facilitating this. The automation requires the development of new localization approaches, as deep open-pit mines are challenging for satellite-based localization systems. Deep funnel-shaped mines reduce the sky-view angle from a certain position onward so that few to no satellites are visible. Therefore, we introduce a new wheel-odometry-aided navigation filter for articulated vehicles that fuses measurements from an inertial measurement unit (IMU), global navigation satellite systems (GNSS), and wheel encoders. Non-holonomic constraints are incorporated by assuming the lateral velocity of each wheel to be zero. We present two different measurement models that either use the wheel encoder signals of the rear wheels or all wheels of the articulated vehicle. This approach enables articulated vehicles to cope with the challenges of open-pit mines. The developed navigation filter is evaluated experimentally with an articulated dumper in two scenarios: A paved parking lot and a gravel pit. With the proposed method, we achieved a mean position error of 0.21 m during a 190 s test drive in the gravel pit with a simulated GNSS interruption of 90 s. This is an improvement of 64 m compared to a state-of-the-art navigation filter that fuses only inertial and GNSS measurements.

Robotic peg-in-hole assembly remains a challenging task due to its high accuracy demand. Previous work tends to simplify the problem by restricting the degree of freedom of the end-effector, or limiting the distance between the target and the initial pose position, which prevents them from being deployed in real-world manufacturing. Thus, we present a Coarse-to-Fine Visual Servoing (CFVS) peg-in-hole method, achieving 6-DoF end-effector motion control based on 3D visual feedback. CFVS can handle arbitrary tilt angles and large initial alignment errors through a fast pose estimation before refinement. Furthermore, by introducing a confidence map to ignore the irrelevant contour of objects, CFVS is robust against noise and can deal with various targets beyond training data. Extensive experiments show CFVS outperforms state-of-the-art methods and obtains 100%, 91%, and 82% average success rates in 3-DoF, 4-DoF, and 6-DoF peg-in-hole, respectively.

Deploying object detectors in embedded devices such as unmanned aerial vehicles (UAVs) comes with many challenges. This is due to both the UAV itself having low embedded resources in terms of computation and memory, and also due to the nature of the captured visual data with the variations in objects' scale, orientation, density, viewpoint, distribution, shape, context and others. It is crucial for the object detector to be robust with high accuracy, real-time with fast inference and light-weight to be applicable. Inspired by YOLO architecture, we propose a novel single-stage detection architecture. Our contributions are, first, feature fusion spatial pyramid pooling (FFSPP) block that applies attention-based feature fusion across both time and space utilizing the information of subsequent frames and scales in an efficient manner. Secondly, we introduce a multi-dilated attention-based cross-stage partial connection (MDACSP) block that helps in increasing the receptive field and producing per-channel modulation weights after aggregating the feature maps across their spatial domain. Third, scaled feature fusion head (SFFH) fuses both the FFSPP block features and the connected MDACSP block features specific for this head. For a more robust result across different scenarios, we perform cross-ensembling with three of the top UAV/traffic surveillance datasets: UAVDT, UA-DETRAC and VisDrone. Our ablation study shows how every contribution improves over the baseline. Our approach yielded the state-of-the-art results in all the aforementioned datasets achieving 89.3% mAP, 93.5% mAP, and 42.9% mAP respectively. Testing the model performance on NVIDIA Jetson Xavier NX board shows a desirable balance between the inference time and the memory cost. We also show qualitatively the model robustness and efficiency across the diverse complex scenarios of these datasets. We hope this work facilitates the advancement of the UAV-based perception in such crucial industrial applications. Show More

Endobronchial intervention is increasingly used as a minimally invasive means for the treatment of pulmonary diseases. In order to reduce the difficulty of manipulation in complex airway networks, robust lumen detection is essential for intraoperative guidance. However, these methods are sensitive to visual artifacts which are inevitable during the surgery. In this work, a cross domain feature interaction (CDFI) network is proposed to extract the structural features of lumens, as well as to provide artifact cues to characterize the visual features. To effectively extract the structural and artifact features, the Quadruple Feature Constraints (QFC) module is designed to constrain the intrinsic connections of samples with various imaging-quality. Furthermore, we design a Guided Feature Fusion (GFF) module to supervise the model for adaptive feature fusion based on different types of artifacts. Results show that the features extracted by the proposed method can preserve the structural information of lumen in the presence of large visual variations, bringing much-improved lumen detection accuracy.

Road traffic scene reconstruction from videos has been desirable by road safety regulators, city planners, researchers, and autonomous driving technology developers. However, it is expensive and unnecessary to cover every mile of the road with cameras mounted on the road infrastructure. This paper presents a method that can process aerial videos to vehicle trajectory data so that a traffic scene can be automatically reconstructed and accurately re-simulated using computers. On average, the vehicle localization error is about 0.1 m to 0.3 m using a consumer-grade drone flying at 120 meters. This project also compiles a dataset of 50 reconstructed road traffic scenes from about 100 hours of aerial videos to enable various downstream traffic analysis applications and facilitate further road traffic related research. The dataset is available at https://github.com/duolu/CAROM.

Robust visual place recognition (VPR) against significant appearance changes is crucial for the life-long operation of mobile robots. Focusing on this task, we propose a Co-Attentive Hierarchical Image Representations (CAHIR) framework for VPR, which unifies attention-sharing global and local descriptor generation into one encoding pipeline. The hierarchical descriptors are applied to a coarse-to-fine VPR system with global retrieval and local geometric verification. To explore high-quality local matches between task-relevant visual elements, a cross-attention mutual enhancement layer is introduced to strengthen the information interaction between the local descriptors. Through the proposed selective matching distillation, the mutual enhancement layer can learn from state-of-the-art local matchers in a distillation manner. After weighted cross-matching of the enhanced local descriptors, geometric verification is applied to evaluate the spatial consistency of the compared image pair. Experiments show CAHIR outperforms the existing global and local representations for VPR in terms of performance and efficiency. Quantitatively, it achieves state-of-the-art results on three city-scale benchmark datasets. Qualitatively, CAHIR proves to attach great importance to task-relevant visual elements and excels at finding local correspondences that are discriminative to the VPR task.

Terrestrial cyborg insects are biohybrid systems integrating living insects as mobile platforms. The insects' locomotion is controlled by the electrical stimulation of their sensory, muscular, or neural systems, in which continuous pulse trains are usually chosen as the stimulation waveform. Although this waveform is easy to generate and can elicit graded responses from the insects, its locomotion control efficiency has not been consistent among existing literature. This study demonstrates an improvement in locomotion control by using a new stimulation protocol, named Burst Stimulation, to stimulate a cyborg beetle's antennae (Zophobas morio). Modulating the continuous pulse train into multiple bursts enhanced the beetle's turning responses. At the same stimulation intensity (amplitude, pulse width, and active duration), the Burst Stimulation improved the turning angle by up to 50% compared to the continuous waveform. Moreover, the beetle's graded response was preserved. Increasing the stimulation frequency from 10 Hz to 40 Hz raised the turning rate by 40 deg/s. In addition, the initial implementation of this protocol in the feedback control-based navigation achieved a success rate of 81%, suggesting its potential use to optimize further the autonomous navigation of terrestrial cyborg insects.

We present the first free-floating autonomous underwater construction system capable of using active bal-lasting to transport cement building blocks efficiently. It is the first free-floating autonomous construction robot to use a paired set of resources: compressed air for buoyancy and a battery for thrusters. In construction trials, our system built structures of up to 12 components and weighing up to 100 Kg (75 Kg in water). Our system achieves this performance by combining a novel one-degree-of-freedom manipulator, a novel two-component cement block construction system that corrects errors in placement, and a simple active ballasting system combined with compliant placement and grasp behaviors. The passive error correcting components of the system minimize the required complexity in sensing and control. We also explore the problem of buoyancy allocation for building structures at scale by defining a convex program which allocates buoyancy to minimize the predicted energy cost for transporting blocks.

Existing multi-agent perception algorithms usually select to share deep neural features extracted from raw sensing data between agents, achieving a trade-off between accuracy and communication bandwidth limit. However, these methods assume all agents have identical neural networks, which might not be practical in the real world. The transmitted features can have a large domain gap when the models differ, leading to a dramatic performance drop in multi-agent perception. In this paper, we propose the first lightweight framework to bridge such domain gaps for multi-agent perception, which can be a plug-in module for most of the existing systems while maintaining confidentiality. Our framework consists of a learnable feature resizer to align features in multiple dimensions and a sparse cross-domain transformer for domain adaption. Extensive experiments on the public multi-agent perception dataset V2XSet have demonstrated that our method can effectively bridge the gap for features from different domains and outperform other baseline methods significantly by at least 8% for point-cloud-based 3D object detection.