ICRA 2022 论文列表

Automation of vineyards cultivation necessitates for mobile robots to retain accurate localization system. The paper introduces a stereo vision-based Graph-Simultaneous Localization and Mapping (Graph-SLAM) pipeline custom-tailored to the specificities of vineyard fields. Graph-SLAM is reinforced with a Loop Closure Detection (LCD) based on semantic segmentation of the vine trees. The Mask R-CNN network is applied to segment the trunk regions of images, on which unique visual features are extracted. These features are used to populate the bag of visual words (BoVW s) retained on the formulated graph. A nearest neighbor search is applied to each query trunk-image to associate each unique feature descriptor with the corresponding node in the graph using a voting procedure. We apply a probabilistic method to select the most suitable loop closing pair and, upon an LCD appearance, the 3D points of the trunks are employed to estimate the loop closure constraint to the graph. The traceable features on trunk segments drastically reduce the number of retained BoVWs, which in turn expedites significantly the loop closure and graph optimization, rendering our method suitable for large scale mapping in vineyards. The pipeline has been evaluated on several data sequences gathered from real vineyards, in different seasons, when the appearance of vine trees vary significantly, and exhibited robust mapping in long distances.

Planning a safe trajectory for an ego vehicle through an environment with occluded regions is a challenging task. Existing methods use some combination of metrics to evaluate a trajectory, either taking a worst case view or allowing for some probabilistic estimate, to eliminate or minimize the risk of collision respectively. Typically, these approaches assume occluded regions of the environment are unsafe and must be avoided, resulting in overly conservative trajectories-particularly when there are no hidden risks present. We propose a local trajectory planning algorithm which generates safe trajectories that maximize observations on un-certain regions. In particular, we seek to gain information on occluded areas that are most likely to pose a risk to the ego vehicle on its future path. Calculating the information gain is a computationally complex problem; our method approximates the maximum information gain and results in vehicle motion that remains safe but is less conservative than state-of-the-art approaches. We evaluate the performance of the proposed method within the CARLA simulator in different scenarios.

Humans have the ability to pour various media, both liquid and granular, to desired ends in various containers. We do this by using multiple senses simultaneously in a constant feedback loop to complete a pouring task. Combining multiple sensing modalities, similar to humans, could aid in robotic pouring control outside of a structured or industrial setting. We present a multi-sensory pouring dataset consisting of human pouring demonstrations of various granular media, coupled with two multi-sensory networks that estimate pouring rate and pouring average height. For both pouring metrics, a combined input of audio and visual data provides a lower median error than either the audio network or visual network. The multi-sensory network achieves a median error of 6.4 mm for average height estimation and 0.06 N/s for pouring rate estimation.

We present a strategy for designing and building very general robot manipulation systems using a general-purpose task-and-motion planner with both engineered and learned modules that estimate properties and affordances of unknown objects. Such systems are closed-loop policies that map from RGB images, depth images, and robot joint encoder measurements to robot joint position commands. We show that this strategy leads to intelligent behaviors even without a priori knowledge regarding the set of objects, their geometries, and their affordances. We show how these modules can be flexibly composed with robot-centric primitives using the PDDLStream task and motion planning framework. Finally, we demonstrate that this strategy can enable a single policy to perform a wide variety of real-world multi-step manipulation tasks, generalizing over a broad class of objects, arrangements, and goals, without prior knowledge of the environment or re-training.

We want robots to complete assigned tasks even when unexpected task pressures arise, either from the robot or the environment. This paper presents a method of both learning sources of task failure in situ and rapidly planning new motions on-the-fly to accommodate them. This “risk-adaptive” approach to robot control uses a few encounters with a novel failure mode to generate a probabilistic failure model which we use to optimize a risk-mitigating motion plan. We demonstrate two toy problems, where risk-adaptive double-integrator agents are introduced to separate environments, each with their own tasks and modes of failure. The agents are not aware a priori of any risks the environments might present, but after one failure, the agents quickly adapt their motion plans and ensure task completion. We further conduct numerical experiments to characterize the algorithm's speed of adaptation with respect to environmental uncertainty. We see this framework as a natural extension for the myriad of robotic applications using model-based motion planners.

This work presents a technique for localization of a smart infrastructure node, consisting of a fisheye camera, in a prior map. These cameras can detect objects that are outside the line of sight of the autonomous vehicles (AV) and send that information to AVs using V2X technology. However, in order for this information to be of any use to the AV, the detected objects should be provided in the reference frame of the prior map that the AV uses for its own navigation. Therefore, it is important to know the accurate pose of the infrastructure camera with respect to the prior map. Here we propose to solve this localization problem in two steps, (i) we perform feature matching between perspective projection of fisheye image and bird's eye view (BEV) satellite imagery from the prior map to estimate an initial camera pose, (ii) we refine the initialization by maximizing the Mutual Information (MI) between intensity of pixel values of fisheye image and reflectivity of 3D LiDAR points in the map data. We validate our method on simulated data and also present results with real world data.

In this paper, we present a data acquisition method to realize a distributed tactile sensor system that can provide wide-range, and high-sensitivity with a small data size for communication. Since the data size is proportional to the number of acquired data and the resolution of the data, we propose systems to increase the resolution of the sensor output values without increasing the amount of data and to reduce the sampling frequency without reducing the time resolution. The system to increase the data resolution is based on a mechanism to dynamically change the gain of an amplifier circuit for a tactile sensor depending on the input value in the local controller. The system to reduce the number of data acquisitions consists of an analog circuit to judge the amount of change in the tactile sensor output based on a threshold at the hardware level. We demonstrate that the system reduces the data size for communication on a sensor system in our robotic hand. The experimental results indicate that the tactile sensor systems could sense high-resolution data with a small data size for communication.

Retrieval-based place recognition is an efficient and effective solution for re-localization within a pre-built map, or global data association for Simultaneous Localization and Mapping (SLAM). The accuracy of such an approach is heavily dependant on the quality of the extracted scene-level representation. While end-to-end solutions - which learn a global descriptor from input point clouds - have demonstrated promising results, such approaches are limited in their ability to enforce desirable properties at the local feature level. In this paper, we introduce a local consistency loss to guide the network towards learning local features which are consistent across revisits, hence leading to more repeatable global descriptors resulting in an overall improvement in 3D place recognition performance. We formulate our approach in an end-to-end trainable architecture called LoGG3D-Net. Experiments on two large-scale public benchmarks (KITTI and MulRan) show that our method achieves mean F1max scores of 0.939 and 0.968 on KITTI and MulRan respectively, achieving state-of-the-art performance while operating in near real-time. The open-source implementation is available at: https://github.com/csiro-robotics/LoGG3D-Net.

We present a lightweight encoder-decoder architecture for monocular depth estimation, specifically designed for embedded platforms. Our main contribution is the Guided Upsampling Block (GUB) for building the decoder of our model. Motivated by the concept of guided image filtering, GUB relies on the image to guide the decoder on upsampling the feature representation and the depth map reconstruction, achieving high resolution results with fine-grained details. Based on multiple GUBs, our model outperforms the related methods on the NYU Depth V2 dataset in terms of accuracy while delivering up to 35.1 fps on the NVIDIA Jetson Nano and up to 144.5 fps on the NVIDIA Xavier NX. Similarly, on the KITTI dataset, inference is possible with up to 23.7 fps on the Jetson Nano and 102.9 fps on the Xavier NX. Our code and models are made publicly available14https://github.com/mic-rud/GuidedDecoding.

Flapping Wing Rotorcraft (FWR) combines flapping and rotating wing motion in one element. Such a hybrid design integrates the high-efficiency characteristics of the rotating wing and the high-lift feature of the flapping wing under low Reynolds number, providing a broader range of simultaneous lift and power efficiency optimization. Nevertheless, the flight performance of the current FWRs is limited by their complex transmission mechanisms. Such mechanical constraints not only induce coupled wing kinematics but also render tedious assembly work and fabrication imperfections. In order to fundamentally address the constraints, we propose a motor-driven FWR with mechanically decoupled wings. The wing of the proposed DFWR is directly actuated by two bi-directional rotating motors instead of using the crank rocker (or alike) transmission. The proposed DFWR flaps within 25Hz to 35Hz, with about 12.4 grams of system weight and 185mm wingspan. With the direct-drive principle, the wing kinematics can be modulated properly by real-time motor control. In particular, we tuned the flapping frequency, stroke amplitude, and mid-stroke angle of the proposed direct-drive FWR to attain its best lift performance. As a result, it can generate about 16 grams of maximum total lift. In order to validate the proposed design, free flight tests have been conducted. The proposed FWR demonstrates stable liftoff.

We present a system for automatic converting of 2D mask object predictions and raw LiDAR point clouds into full 3D bounding boxes of objects. Because the LiDAR point clouds are partial, directly fitting bounding boxes to the point clouds is meaningless. Instead, we suggest that obtaining good results requires sharing information between all objects in the dataset jointly, over multiple frames. We then make three improvements to the baseline. First, we address ambiguities in predicting the object rotations via direct optimization in this space while still backpropagating rotation prediction through the model. Second, we explicitly model outliers and task the network with learning their typical patterns, thus better discounting them. Third, we enforce temporal consistency when video data is available. With these contributions, our method significantly outperforms previous work despite the fact that those methods use significantly more complex pipelines, 3D models and additional human-annotated external sources of prior information.

Grasping objects whose physical properties are unknown is still a great challenge in robotics. Most solutions rely entirely on visual data to plan the best grasping strategy. However, to match human abilities and be able to reliably pick and hold unknown objects, the integration of an artificial sense of touch in robotic systems is pivotal. This paper describes a novel model-based slip detection pipeline that can predict possibly failing grasps in real-time and signal a necessary increase in grip force. As such, the slip detector does not rely on manually collected data, but exploits physics to generalize across different tasks. To evaluate the approach, a state-of-the-art vision-based tactile sensor that accurately estimates distributed forces was integrated into a grasping setup composed of a six degrees-of-freedom cobot and a two-finger gripper. Results show that the system can reliably predict slip while manipulating objects of different shapes, materials, and weights. The sensor can detect both translational and rotational slip in various scenarios, making it suitable to improve the stability of a grasp.

Multi-agent interactions are important to model for forecasting other agents' behaviors and trajectories. At a certain time, to forecast a reasonable future trajectory, each agent needs to pay attention to the interactions with only a small group of most relevant agents instead of unnecessarily paying attention to all the other agents. However, existing attention modeling works ignore that human attention in driving does not change rapidly, and may introduce fluctuating attention across time steps. In this paper, we formulate an attention model for multi-agent interactions based on a total variation temporal smoothness prior and propose a trajectory prediction architecture that leverages the knowledge of these attended interactions. We demonstrate how the total variation attention prior along with the new sequence prediction loss terms leads to smoother attention and more sample-efficient learning of multi-agent trajectory prediction, and show its advantages in terms of prediction accuracy by comparing it with the state-of-the-art approaches on both synthetic and naturalistic driving data. We demonstrate the performance of our algorithm for trajectory prediction on the INTERACTION dataset on our website11https://sites.google.com/view/smoothness-attention.

As robots gain capabilities to enter our humancentric world, they require formalism and algorithms that enable smart and efficient interactions. This is challenging, especially for robotic manipulators with complex tasks that may require collaboration with humans. Prior works approach this problem through reactive synthesis and generate strategies for the robot that guarantee task completion by assuming an adversarial human. While this assumption gives a sound solution, it leads to an “unfriendly” robot that is agnostic to the human intentions. We relax this assumption by formulating the problem using the notion of regret. We identify an appropriate definition for regret and develop regret-minimizing synthesis framework that enables the robot to seek cooperation when possible while preserving task completion guarantees. We illus-trate the efficacy of our framework via various case studies.

Robot-mediated interventions are one promising and novel approach for encouraging motor exploration in young children, but knowledge about the effectiveness of toy-like features for child-robot interaction is limited. We were interested in understanding the characteristics of current toys to inform the design of interactive abilities for assistive robots. This work first provides a systematic review of toy characteristics in n=154n=154 Fisher-Price products and then analyzes the effectiveness of common and uncommon toy-like behaviors from our custom assistive robot. Toy review results showed that light and sound features were significantly more common than bubbles, wheels, and self-propulsion. Exploratory play sessions with our assistive robot showed that bubbles were significantly more successful at encouraging child motion than other robot behaviors. Further, all studied robot behaviors demonstrated the capability to encourage child motion. The products of this work can inform the efforts of human-robot interaction and child development experts who study child mobility interventions.

Legged robots are physically capable of traversing a wide range of challenging environments, but designing controllers that are sufficiently robust to handle this diversity has been a long-standing challenge in robotics. Reinforcement learning presents an appealing approach for automating the controller design process and has been able to produce remarkably robust controllers when trained in a suitable range of environments. However, it is difficult to predict all likely conditions the robot will encounter during deployment and enumerate them at training-time. What if instead of training controllers that are robust enough to handle any eventuality, we enable the robot to continually learn in any setting it finds itself in? This kind of real-world reinforcement learning poses a number of challenges, including efficiency, safety, and autonomy. To address these challenges, we propose a practical robot reinforcement learning system for fine-tuning locomotion policies in the real world. We demonstrate that a modest amount of real-world training can substantially improve performance during deployment, and this enables a real A1 quadrupedal robot to autonomously fine-tune multiple locomotion skills in a range of environments, including an outdoor lawn and a variety of indoor terrains. (Videos and code11https://sites.google.com/berkele.edu/fine-tuning-locomotion)

In our daily life, there are many objects represented by cylindrical shapes and ellipsoids. The tops of these objects are formed by elliptic shape primitives. Thus, it is available for a robot to manipulate these objects by ellipse detection. In this work, we propose a novel approach to generating ground truth for training the model based on domain randomization. Using synthetic data generated in this manner, we build an end-to-end deep neural network with a detection backbone and then, combine multiple branches archived from the backbone for sharing the multiple-scale features; further, after employing active rotation filters, the features pass through the region proposal net to form the prediction branches of the box, orientation regression, and object classification; finally, these branches are fused to do ellipse detection, allowing robotic manipulations of cylinders and ellipsoids. To demonstrate the capabilities of the proposed detector, we show the comparison results with the state-of-the-art detector on synthetic and public datasets. The proposed model for ellipse detection and data generation pipeline based on domain randomization in a simulation are evaluated by a series of robotic manipulations implemented in real application scenarios. The results illustrate a high success rate on real-world grasp attempts despite having only been trained on a synthetic dataset. (A video of some robotic experiments is available on YouTube: https://youtu.be/Ueg1XSI2S98).

Optimal motion planning is a long-studied problem with a wide range of applications in robotics, from grasping to navigation. While sampling-based motion planning methods have made solving such problems significantly more feasible, these methods still often struggle in high-dimensional spaces wherein exploration is computationally costly. In this paper, we propose a new motion planning algorithm that reduces the computational burden of the exploration process. The proposed algorithm utilizes a guidance policy acquired offline through model-free reinforcement learning. The guidance policy is used to bias the exploration process in motion planning and to guide it toward promising regions of the state space. Moreover, we show that the gradients of the corresponding learned value function can be used to locally fine-tune the sampled states. We empirically demonstrate that the proposed approach can significantly reduce planning time and improve success rate and path quality.

Vision-based tactile sensors typically utilize a deformable elastomer and a camera mounted above to provide high-resolution image observations of contacts. Obtaining accurate volumetric meshes for the deformed elastomer can provide direct contact information and benefit robotic grasping and manipulation. This paper focuses on learning to synthesize the volumetric mesh of the elastomer based on the image imprints acquired from vision-based tactile sensors. Synthetic image-mesh pairs and real-world images are gathered from 3D finite element methods (FEM) and physical sensors, respectively. A graph neural network (GNN) is introduced to learn the image-to-mesh mappings with supervised learning. A self-supervised adaptation method and image augmentation techniques are proposed to transfer networks from simulation to reality, from primitive contacts to unseen contacts, and from one sensor to another. Using these learned and adapted networks, our proposed method can accurately reconstruct the deformation of the real-world tactile sensor elastomer in various domains, as indicated by the quantitative and qualitative results.