ICRA 2023 论文列表

In this paper, we address the lack of datasets for – and the issue of reproducibility in – collaborative SLAM pose graph optimizers by providing a novel pose graph generator. Our pose graph generator, kollagen, is based on a random walk in a planar grid world, similar to the popular M3500 dataset for single agent SLAM. It is simple to use and the user can set several parameters, e.g., the number of agents, the number of nodes, loop closure generation probabilities, and standard deviations of the measurement noise. Furthermore, a qualitative execution time analysis of our pose graph generator showcases the speed of the generator in the tunable parameters. In addition to the pose graph generator, our paper provides two example datasets that researchers can use out-of-the-box to evaluate their algorithms. One of the datasets has 8 agents, each with 3500 nodes, and 67645 constraints in the pose graphs, while the other has 5 agents, each with 10000 nodes, and 76134 constraints. In addition, we show that current state-of-the-art pose graph optimizers are able to process our generated datasets and perform pose graph optimization. The data generator can be found at https://github.com/EricssonResearch/kollagen.

We present a novel real-time dense and semantic neural field mapping system that uses only monocular images as input. Our scene representation is a dense continuous radiance field represented by a Multi-Layer Perceptron (MLP), trained from scratch in real-time. We build on high-performance sparse visual SLAM and use camera poses and sparse keypoint depths as supervision alongside RGB keyframes. Since no prior training is required, our system flexibly fits to arbitrary scale and structure at runtime, and works even with strong specular reflections. We demonstrate reconstruction over a range of scenes from small indoor to large outdoor spaces. We also show that the method can straightforwardly benefit from additional inputs such as learned depth priors or semantic labels for more precise and advanced mapping.

Zero-shot object detection has shown its ability to overcome the problems of data scarcity and novel classes. Existing methods generally utilize static semantic vectors to classify objects and guide the network to map visual features to semantic vectors. However, the distribution of semantic vectors cannot adequately represent visual features, which makes migration from seen to unseen classes difficult. This work explores the dynamic semantic vector method to align the distributions of semantic vectors and visual features. The main challenge is to get a more reasonable distribution of semantic vectors. To address this issue, we proposed a two-way classification branch network and introduce N-pair loss into the dynamic semantic vector optimization process. Experiments on the MS-COCO dataset and SiTi (a real-world autonomous driving dataset collected by us) demonstrate the effectiveness and generalization of our method. Our code is available at https://github.com/HaoyuLizju/ZSD_tcb

In this paper, we focus on the problem of efficiently locating a target object described with free-form text using a mobile robot equipped with vision sensors (e.g., an RGBD camera). Conventional active visual search predefines a set of objects to search for, rendering these techniques restrictive in practice. To provide added flexibility in active visual searching, we propose a system where a user can enter target commands using free-form text; we call this system Zero-shot Active Visual Search (ZAVIS). ZAVIS detects and plans to search for a target object inputted by a user through a semantic grid map represented by static landmarks (e.g., desk or bed). For efficient planning of object search patterns, ZAVIS considers commonsense knowledge-based co-occurrence and predictive uncertainty while deciding which landmarks to visit first. We validate the proposed method with respect to SR (success rate) and SPL (success weighted by path length) in both simulated and real-world environments. The proposed method outperforms previous methods in terms of SPL in simulated scenarios, and we further demonstrate ZAVIS with a Pioneer-3AT robot in real-world studies.

Humans naturally exploit haptic feedback during contact-rich tasks like loading a dishwasher or stocking a bookshelf. Current robotic systems focus on avoiding unexpected contact, often relying on strategically placed environment sensors. Recently, contact-exploiting manipulation policies have been trained in simulation and deployed on real robots. However, they require some form of real-world adaptation to bridge the sim-to-real gap, which might not be feasible in all scenarios. In this paper we train a contact-exploiting manipulation policy in simulation for the contact-rich household task of loading plates into a slotted holder, which transfers without any fine-tuning to the real robot. We investigate various factors necessary for this zero-shot transfer, like time delay modeling, memory representation, and domain randomization. Our policy transfers with minimal sim-to-real gap and significantly outperforms heuristic and learnt baselines. It also generalizes well to a cup and plates of different sizes and weights. The project website is https://sites.google.com/view/compliant-object-insertion.

Humans are capable of abstracting various tasks as different combinations of multiple attributes. This perspective of compositionality is vital for human rapid learning and adaption since previous experiences from related tasks can be combined to generalize across novel compositional settings. In this work, we aim to achieve zero-shot policy generalization of Reinforcement Learning (RL) agents by leveraging the task compositionality. Our proposed method is a meta-RL algorithm with disentangled task representation, explicitly encoding different aspects of the tasks. Policy generalization is then performed by inferring unseen compositional task representations via the obtained disentanglement without extra exploration. The evaluation is conducted on three simulated tasks and a challenging real-world robotic insertion task. Experimental results demonstrate that our proposed method achieves policy generalization to unseen compositional tasks in a zero-shot manner.

Object goal visual navigation is a challenging task that aims to guide a robot to find the target object based on its visual observation, and the target is limited to the classes pre-defined in the training stage. However, in real households, there may exist numerous target classes that the robot needs to deal with, and it is hard for all of these classes to be contained in the training stage. To address this challenge, we study the zero-shot object goal visual navigation task, which aims at guiding robots to find targets belonging to novel classes without any training samples. To this end, we also propose a novel zero-shot object navigation framework called semantic similarity network (SSNet). Our framework use the detection results and the cosine similarity between semantic word embeddings as input. Such type of input data has a weak correlation with classes and thus our framework has the ability to generalize the policy to novel classes. Extensive experiments on the AI2-THOR platform show that our model outperforms the baseline models in the zero-shot object navigation task, which proves the generalization ability of our model. Our code is available at: https://github.com/pioneer-innovation/Zero-Shot-Object-Navigation.

In the era of deep learning, data is the critical determining factor in the performance of neural network models. Generating large datasets suffers from various challenges such as scalability, cost efficiency and photorealism. To avoid expensive and strenuous dataset collection and annotations, researchers have inclined towards computer-generated datasets. However, a lack of photorealism and a limited amount of computer-aided data has bounded the accuracy of network predictions. To this end, we present WorldGen - an open source framework to automatically generate countless structured and unstructured 3D photorealistic scenes such as city view, object collection, and object fragmentation along with its rich ground truth annotation data. WorldGen being a generative model gives the user full access and control to features such as texture, object structure, motion, camera and lens properties for better generalizability by diminishing the data bias in the network. We demonstrate the effectiveness of WorldGen by evaluating deep optical flow. We hope such a tool can open doors for future research in a myriad of domains related to robotics and computer vision by reducing manual labor and cost for acquiring rich and high-quality data.

Electrostatic actuators provide a promising approach to creating soft robotic sheets, due to their flexible form factor, modular integration, and fast response speed. However, their control requires kilo-Volt signals and understanding of complex dynamics resulting from force interactions by on-board and environmental effects. In this work, we demonstrate an untethered planar five-actuator piezoelectric robot powered by batteries and on-board high-voltage circuitry, and controlled through a wireless link. The scalable fabrication approach is based on bonding different functional layers on top of each other (steel foil substrate, actuators, flexible electronics). The robot exhibits a range of controllable motions, including bidirectional crawling (up to ~0.6 cm/s), turning, and in-place rotation (at ~1 degree/s). High-speed videos and control experiments show that the richness of the motion results from the interaction of an asymmetric mass distribution in the robot and the associated dependence of the dynamics on the driving frequency of the piezoelectrics. The robot's speed can reach 6 cm/s with specific payload distribution.

Many existing datasets for lidar place recognition are solely representative of structured urban environments, and have recently been saturated in performance by deep learning based approaches. Natural and unstructured environments present many additional challenges for the tasks of long-term localisation but these environments are not represented in currently available datasets. To address this we introduce Wild-Places, a challenging large-scale dataset for lidar place recognition in unstructured, natural environments. Wild-Places contains eight lidar sequences collected with a handheld sensor payload over the course of fourteen months, containing a total of 63K undistorted lidar submaps along with accurate 6DoF ground truth. This dataset contains multi-ple revisits both within and between sequences, allowing for both intra-sequence (i.e., loop closure detection) and inter-sequence (i.e., re-localisation) tasks. We also benchmark several state-of-the-art approaches to demonstrate the challenges that this dataset introduces, particularly the case of long-term place recognition due to natural environments changing over time. Our dataset and code is available at https://csiro-robotics.github.io/Wild-Places

Cross-view geolocalization, a supplement or replacement for GPS, localizes an agent within a search area by matching images taken from a ground-view camera to overhead images taken from satellites or aircraft. Although the viewpoint disparity between ground and overhead images makes crossview geolocalization challenging, significant progress has been made assuming that the ground agent has access to a panoramic camera. For example, our prior work (WAG) introduced changes in search area discretization, training loss, and particle filter weighting that enabled city-scale panoramic cross-view geolocalization. However, panoramic cameras are not widely used in existing robotic platforms due to their complexity and cost. Non-panoramic cross-view geolocalization is more applicable for robotics, but is also more challenging. This paper presents Restricted FOV Wide-Area Geolocalization (ReWAG), a cross-view geolocalization approach that generalizes WAG for use with standard, non-panoramic ground cameras by creating pose-aware embeddings and providing a strategy to incorporate particle pose into the Siamese network. ReWAG is a neural network and particle filter system that is able to globally localize a mobile agent in a GPS-denied environment with only odometry and a 90° FOV camera, achieving similar localization accuracy as what WAG achieved with a panoramic camera and improving localization accuracy by a factor of 100 compared to a baseline vision transformer (ViT) approach.

In this paper we propose a novel algorithm, Wi-Closure, to improve the computational efficiency and robustness of loop closure detection in multi-robot SLAM. Our approach decreases the computational overhead of classical approaches by pruning the search space of potential loop closures, prior to evaluation by a typical multi-robot SLAM pipeline. Wi-Closure achieves this by identifying candidates that are spatially close to each other measured via sensing over the wireless communication signal between robots, even when they are operating in non-line-of-sight or in remote areas of the environment from one another. We demonstrate the validity of our approach in simulation and in hardware experiments. Our results show that using Wi-closure greatly reduces computation time, by 54.1% in simulation and 76.8% in hardware experiments, compared with a multi-robot SLAM baseline. Importantly, this is achieved without sacrificing accuracy. Using Wi-closure reduces absolute trajectory estimation error by 98.0% in simulation and 89.2% in hardware experiments. This improvement is partly due to Wi-Closure's ability to avoid catastrophic optimization failure that typically occurs with classical approaches in challenging repetitive environments.

Recently, Model Predictive Contouring Control (MPCC) has arisen as the state-of-the-art approach for model-based agile flight. MPCC benefits from great flexibility in trading-off between progress maximization and path following at runtime without relying on globally optimized trajectories. However, finding the optimal set of tuning parameters for MPCC is challenging because (i) the full quadrotor dynamics are non-linear, (ii) the cost function is highly non-convex, and (iii) of the high dimensionality of the hyperparameter space. This paper leverages a probabilistic Policy Search method—Weighted Maximum Likelihood (WML)—to automatically learn the optimal objective for MPCC. WML is sample-efficient due to its closed-form solution for updating the learning parameters. Additionally, the data efficiency provided by the use of a model-based approach allows us to directly train in a high-fidelity simulator, which in turn makes our approach able to transfer zero-shot to the real world. We validate our approach in the real world, where we show that our method outperforms both the previous manually tuned controller and the state-of-the-art auto-tuning baseline reaching speeds of 75 km/h.

Weakly supervised referring expression grounding aims to train a model without the manual labels between image regions and referring expressions during the training phase. Current predominant models often adopt deep structures to reconstruct the region-expression correspondence. A crucial deficiency of the existing approaches lies in that these models neglect to exploit potential valuable information to further improve their grounding performance. To address this issue, we leverage knowledge distillation as a unique scheme to excavate and transfer helpful information for acquiring a better model. Specifically, we propose a target-guided knowledge distillation framework that accounts for region-expression pairs reconstruction and matching. We reactivate the target-related prediction information learned by a pre-trained teacher model and transfer the target-related prediction knowledge from the teacher to guide the training process and boost the performance of the student model. We conduct extensive experiments on three benchmark datasets, i.e., RefCOCO, RefCOCO+, and RefCOCOg. Without bells and whistles, our approach achieves state-of-the-art results on several splits of benchmark datasets. The implementation codes and trained models are available at: https://github.com/dami23/WREG_KD.

Motion forecasting for autonomous driving is a challenging task because complex driving scenarios involve a heterogeneous mix of static and dynamic inputs. It is an open problem how best to represent and fuse information about road geometry, lane connectivity, time-varying traffic light state, and history of a dynamic set of agents and their interactions into an effective encoding. To model this diverse set of input features, many approaches proposed to design an equally complex system with a diverse set of modality specific modules. This results in systems that are difficult to scale, extend, or tune in rigorous ways to trade off quality and efficiency. In this paper, we present Wayformer, a family of simple and homogeneous attention based architectures for motion forecasting. Wayformer offers a compact model description consisting of an attention based scene encoder and a decoder. In the scene encoder we study the choice of early, late and hierarchical fusion of input modalities. For each fusion type we explore strategies to trade off efficiency and quality via factorized attention or latent query attention. We show that early fusion, despite its simplicity, is not only modality agnostic but also achieves state-of-the-art results on both Waymo Open Motion Dataset (WOMD) and Argoverse leaderboards, demonstrating the effectiveness of our design philosophy.

Compared to 2D lanes, real 3D lane data is difficult to collect accurately. In this paper, we propose a novel method for training 3D lanes with only 2D lane labels, called weakly supervised 3D lane detection WS-3D-Lane. By assumptions of constant lane width and equal height on adjacent lanes, we indirectly supervise 3D lane heights in the training. To overcome the problem of the dynamic change of the camera pitch during data collection, a camera pitch self-calibration method is proposed. In anchor representation, we propose a double-layer anchor with non-maximum suppression (NMS) method, which enables the anchor-based method to predict two lane lines that are close. Experiments are conducted on the base of 3D-LaneNet under two supervision methods. Under weakly supervised setting, our WS-3D-Lane outperforms previous 3D-LaneN et: F-score rises to 92.3% on Apollo 3D synthetic dataset, and F1 rises to 74.5% on ONCE-3DLanes. Meanwhile, WS-3D-Lane in purely supervised setting makes more increments and outperforms state-of-the-art. To the best of our knowledge, WS-3D- Lane is the first try of 3D lane detection under weakly supervised setting. Our code is available on https://github.com/SAIC-Vision/WS-3D-Lane.

Domain generalization for semantic segmentation is highly demanded in real applications, where a trained model is expected to work well in previously unseen domains. One challenge lies in the lack of data which could cover the diverse distributions of the possible unseen domains for training. In this paper, we propose a WEb-image assisted Domain GEneralization (WEDGE) scheme, which is the first to exploit the diversity of web-crawled images for generalizable semantic segmentation. To explore and exploit the real-world data distributions, we collect web-crawled images which present large diversity in terms of weather conditions, sites, lighting, camera styles, etc. We also present a method which injects styles of the web-crawled images into training images on-the-fly during training, which enables the network to experience images of diverse styles with reliable labels for effective training. Moreover, we use the web-crawled images with their predicted pseudo labels for training to further enhance the capability of the network. Extensive experiments demonstrate that our method clearly outperforms existing domain generalization techniques.

Filter-based shared control aims to accept and augment an operator's ability to control a robot. Current solutions accept actions based on their direction aligning with the robot's optimal policy. These strategies reject a human's small corrective actions if they conflict with the robot's direction and accept too aggressive actions as long as they are consistent with the robot's direction. Such strategies may cause task failures and the operator's feeling of loss of control. To close the gap, we propose WE-Filter, which has flexible, adaptive criteria allowing the operator's small corrective actions and tempering too aggressive ones. Inspired by classical work-energy impact problems between two dynamic, interactive bodies, both inputs' properties (direction and magnitude) are inherently considered, creating intuitive, adaptive bounds to accept sensible actions. The model identifies behaviors before and after impact. The rationale is that each timestep of shared control acts as an impact between the operator's and the robot's policies, where post-impact behaviors depend on their previous behaviors. As time continues, a series of impacts occur. The aim is to minimize impacts that occur to reach an agreement faster and reduce strong reactionary behaviors. Our model determines flexible acceptance criteria to bound a mismatch of magnitude and finds a replacement action for conflicting policies. The WE-Filter achieves better task performance, the ratio of accepted actions, and action similarity than the existing methods.

This paper introduces a novel approach to GPS-denied visual navigation of a robot team over a wide (i.e., out of line of sight) area which we call WAVN (Wide Area Visual Navigation). Application domains include small-scale precision agriculture as well as exploration and surveillance. The proposed approach requires no exploration or map generation, merging, and updating, some of the most computationally intensive aspects of multi-robot navigation, especially in dynamic environments and for long-term deployments. In contrast, we extend the visual homing paradigm to leverage visual information from the entire team to allow a robot to home to a distant location. Since it only employs the latest imagery, the approach can be resilient to the current state of the environment. WAVN requires three components: identification of common landmarks between robots, a communication infrastructure, and an algorithm to find a sequence of common landmarks to navigate to a goal. The principal contribution of this paper is the navigation algorithm in addition to simulation and physical robot results characterizing performance. The approach is also compared to more traditional map-based approaches.