IROS 2021 Papers

The key role of tactile sensing for human grasping and manipulation is widely acknowledged, but most industrial robot grippers and even multi-fingered hands are still designed and used without any tactile sensors. While the basic design principles for resistive or capacitive sensors are well known, several factors keep tactile sensing from large-scale deployment — high sensor costs, short lifespan, poor reliability, difficult production processes, a lack of suitable software and tools for system integration, and the unique requirement for tactile sensors to conform to application-specific shapes.In this work, we describe a very simple but efficient approach to design low-cost resistive matrix sensors, where sensor layout and geometry, taxel-size, and measurement sensitivity can be customized over a wide range. Sensor assembly needs nothing more than a hobby cutting plotter for precise cutting of aluminum tape and Velostat foils, as well as adhesive plastic tape. Our electronics combines transimpedance amplifiers with common Arduino microcontrollers, supporting standard communication protocols, and using either cabled or wireless data transfer to the host. We present three different application examples and sketch our ROS software for sensor calibration and visualization. All parts of our project, including detailed building instructions, bill-of-materials, electronics, and firmware are available open-source.

In this work we propose a learning approach to high-precision robotic assembly problems. We focus on the contact-rich phase, where the assembly pieces are in close contact with each other. Unlike many learning-based approaches that heavily rely on vision or spatial tracking, our approach takes force/torque in task space as the only observation. Our training environment is robotless, as the end-effector is not attached to any specific robot. Trained policies can then be applied to different robotic arms without re-training. This approach can greatly reduce complexity to perform contact-rich robotic assembly in the real world, especially in unstructured settings such as in architectural construction. To achieve it, we have developed a new distributed RL agent, named Recurrent Distributed DDPG (RD2), which extends Ape-X DDPG[1] with recurrency and makes two structural improvements on prioritized experience replay[2]. Our results show that RD2 is able to solve two fundamental high-precision assembly tasks, lap-joint and peg-in-hole, and outperforms two state-of-the-art algorithms, Ape-X DDPG and PPO with LSTM. We have successfully evaluated our robot-agnostic policies on three robotic arms, Kuka KR60, Franka Panda, and UR10, in simulation. The video presenting our experiments is available at https://sites.google.com/view/rd2-rl

Not only on earth, but also in space, robot systems are increasingly becoming essential elements in our lives such as a mobile exploration robot on Mars. Satellites are also an indispensable field in space robot systems, for example, from low-orbit satellites for self-driving vehicles to small satellites launched for various purposes. A lot of research is being conducted on an automated system using more and more satellites. In particular, earth observation using satellite images is being used in various fields such as disaster prediction, damage analysis, and land cover classification. There are three main types of satellite imagery used in automation systems: optical, Synthetic Aperture Radar (SAR), and infrared. Unlike optical satellite, which is heavily influenced by weather and light, SAR satellite can acquire images in all-weather conditions. Thanks to this advantage, SAR satellite images are used in many fields, in particular, water segmentation. There are various traditional SAR image-based water segmentation methods based on thresholding technique. However, these methods are not suitable for rapidly processing a large amount of SAR images because they require a manual operation to set different thresholds for each image. In this paper, we create a large-scale dataset for water segmentation of KOrean Multi-Purpose SATellite (KOMPSAT-5) containing more than 3,000 images. We perform water segmentation using representative deep learning-based segmentation models such as Fully Convolutional Networks (FCN), U-Net, DeepUNet, and High Resolution Network (HRNet). Experimental results show that high performance of water segmentation can be obtained when a large number of training images are used for all five segmentation models. In addition, we confirm the possibility of the automatic water segmentation system from a large amount of SAR images, away from traditional manual work.

In various control task domains, existing controllers provide a baseline level of performance that—though possibly suboptimal—should be maintained. Reinforcement learning (RL) algorithms that rely on extensive exploration of the state and action space can be used to optimize a control policy. However, fully exploratory RL algorithms may decrease performance below a baseline level during training. In this paper, we address the issue of online optimization of a control policy while minimizing regret with respect to a baseline policy performance. We present a joint imitation-reinforcement learning framework, denoted JIRL. The learning process in JIRL assumes the availability of a baseline policy and is designed with two objectives in mind (a) training while leveraging demonstrations from the baseline policy to minimize regret with respect to the baseline policy, and (b) eventually surpassing the baseline performance. JIRL addresses these objectives by initially learning to imitate the baseline policy and gradually shifting control from the baseline to an RL agent. Experimental results show that JIRL effectively accomplishes the aforementioned objectives in several, continuous action-space domains. The results demonstrate that JIRL is comparable to a state-of-the-art algorithm in its final performance while incurring significantly lower baseline regret during training. Moreover, the results show a reduction factor of up to 21 in baseline regret over a trust-region-based approach that guarantees monotonic policy improvement.

Concentric Tube Robots (CTR) have been gaining ground in minimally-invasive robotic surgeries due to their small footprint, compliance, and high dexterity. CTRs can assure safe interaction with soft tissue, provided that precise and effective motion control is achieved. Controlling the motion of CTRs is still challenging. Commonly used model-based control approaches often employ simplified geometric/dynamic assumptions, which could be very inaccurate in the presence of unmodelled disturbances and external interaction forces. Additionally, application of emerging data-driven algorithms in real-time control of CTRs is limited due to the fact that these controllers require considerable amount of time to let the algorithm develop enough to reach a desired accuracy and relevancy. In this paper, we present a hybrid approach to overcome the aforementioned difficulties. This hybrid solution uses the solution of a kinematic model of the robot to estimate initial values for a model-free data-driven method. The proposed algorithm combines both model-based and data-driven algorithms to provide real-time motion control of CTRs interacting with an unknown external environment. Three different simulations studies were performed to thoroughly evaluate the efficacy of the proposed hybrid control approach as compared to two common model-based and data-driven control techniques. The results demonstrate superior performance of the proposed method. The root-mean-square error of the proposed hybrid approach is less than 1.1 mm, which is 9 times less than a common model-based controller.

Spherical robots are typically comprised of an actuation unit enclosed by a spherical shell. Among nonholonomic systems, spherical robots offer the best maneuverability and lowest energy consumption (due to their omnidirectional movement and single contact point with the ground). This allows them to traverse rough and uneven terrains. Further, using their ability to roll on the ground, they can provide a significantly higher operating time compared to aerial-only robots. Unfortunately, these robots are under-emphasized by researchers compared to other robots (i.e., legged or wheeled robots). Additionally, despite their potential to be used in a multitude of real-world applications, spherical robots have not been successfully adopted by the industry. This is due to the lack of controllability and traversability of the developed designs. In this paper, we introduce a hybrid rolling/flying robot. This design benefits from a flywheel to reduce the effects of the terrain (shocks and vibrations) on the camera and sensors. Our design allows the application of existing control algorithms of drones (such as PX4) on a rolling system. In addition, we propose a dynamics model that can use the point cloud representation of the terrain to simulate the motion of the system with applications in real-time modeling and control.

Vehicle dynamic models are the key to bridge the gap between simulation and real road test in autonomous driving. An accurate vehicle model allows control algorithms in simulation being transferred to real road test with same quality. In this paper, we present a dynamic model residual correction framework (DRF) for vehicle dynamic modeling. DRF provides a general accuracy improvement framework on existing vehicle dynamic models. On top of any existing open-loop dynamic model, this framework builds a Residual Correction Model (RCM) by integrating deep Neural Networks (NN) with Stochastic Variational Gaussian Process (SVGP) model. RCM takes a sequence of vehicle control commands and dynamic states for a certain time duration as modeling inputs, extracts underlying context from this sequence with deep encoder networks, and predicts open-loop dynamic model prediction errors. Five vehicle dynamic models are derived from DRF via encoder variations. Our contribution is consolidated with evaluation of the absolute trajectory error and the similarity between DRF outputs and the ground truth. Compared to classic rule-based and learning-based vehicle dynamic models, DRF accomplishes as high as 74.12% to 85.02% of the absolute trajectory error drop among all DRF variations.

Lenticular angle gauge (LEAG) is a planar pattern that visualizes relative attitude by the position of the black line, which moves according to the viewing angle. The authors developed a new LEAG in which the direction of movement of the black line is 90◦ different from the previous LEAG, and used it to develop a non-square high-accuracy fiducial marker. The new marker realized accurate pose estimation with a position error of 0.15% of the distance and an attitude error of 0.5°. This research contributes to the development of high-accuracy markers with a more flexible design. This paper describes the principle and behavior of the new LEAG, the design of the new fiducial marker, and the pose estimation algorithm, followed by the results of performance verification experiments.

Quadruped locomotion is a challenging task for learning-based algorithms. It requires tedious manual tuning and is difficult to deploy in reality due to the reality gap. In this paper, we propose a quadruped robot learning system for agile locomotion which does not require any pre-training and works well in various real-world terrains. We introduce a hierarchical learning framework that uses reinforcement learning as the high-level policy to adjust the low-level trajectory generator for better adaptability to the terrain. We compact the observation and action space of the reinforcement learning to deploy it on a host computer in reality. Besides, we design a trajectory generator guided by robot posture, which can generate adaptive foot trajectory to interact with the environment. Experimental results show that our system can be easily deployed in reality while only trained in simulation, and also has the advantages of fast convergence and good terrain adaptability. The supplementary video demonstration is available at https://vsislab.github.io/hfql/.

Many manipulation tasks combine high-level discrete planning over actions with low-level motion planning over continuous robot motions. Task and motion planning (TMP) provides a powerful general framework to combine discrete and geometric reasoning, and solvers have been previously proposed for single-robot problems. Multi-robot TMP expands the range of TMP problems that can be solved but poses significant challenges when considering scalability and solution quality. We present a general TMP framework designed for multiple robotic manipulators. This is based on two contributions. First, we propose an optimal task planner designed to support simultaneous discrete actions. Second, we introduce an intermediate scheduler layer between task planner and motion planner to evaluate alternate robot assignments to these actions. This aggressively explores the search space and typically reduces the number of expensive task planning calls. Several benchmarks with a rich set of actions for two manipulators are evaluated. We show promising results in scalability and solution quality of our TMP framework with the scheduler for up to six objects. A demonstration indicates scalability to up to five robots.

The environment of most real-world scenarios such as malls and supermarkets changes at all times. A pre-built map that does not account for these changes becomes out-of-date easily. Therefore, it is necessary to have an up-to-date model of the environment to facilitate long-term operation of a robot. To this end, this paper presents a general lifelong simultaneous localization and mapping (SLAM) framework. Our framework uses a multiple session map representation, and exploits an efficient map updating strategy that includes map building, pose graph refinement and sparsification. To mitigate the unbounded increase of memory usage, we propose a map-trimming method based on the Chow-Liu maximum-mutual-information spanning tree. The proposed SLAM framework has been comprehensively validated by over a month of robot deployment in real supermarket environment. Furthermore, we release the dataset collected from the indoor and outdoor changing environment with the hope to accelerate lifelong SLAM research in the community. Our dataset is available at https://github.com/sanduan168/lifelong-SLAM-dataset.

Deep learning algorithms such as Convolutional Neural Networks (CNNs) are currently used to solve a range of robotics and computer vision problems. These networks typically estimate the desired representation in a single forward pass and must therefore learn to converge from a wide range of initial conditions to a precise result. This is challenging, and has led to increased interest in the development of separate refinement modules which learn to improve a given initial estimate, thus reducing the required search space. Such modules are usually developed ad-hoc for each given application, often requiring significant engineering investment. In this work we propose a generic innovation-based CNN. Our CNN is implemented along with a stochastic gradient descent (SGD) algorithm to iteratively refine a given initial estimate. The proposed approach provides a general framework for the development of refinement modules applicable to a wide range of robotics problems. We apply this framework to object pose estimation and depth estimation and demonstrate significant improvement over the initial estimates, in the range of 4.2 - 8.1%, for both applications.

This study proposes a technique for determining completion of robotic tasks considering a characteristics that some features on force responses are common among various objects. In particular, this paper focuses on the click response because its pattern is common to many objects, although the magnitude of the response differs depending on the object. A discriminator for detecting the click response based on the force sensor values, which combines mel-frequency cepstral coefficient and time-delay neural network is introduced. Based on this discriminator, how to improve the generalization performance to untrained objects is investigated to detect click responses in general for various objects. The experimental results show that the performance can be improved by including several kinds of objects with close time constants and different amplitudes of click responses in the training data.

We consider the Stochastic Orienteering Problem with random traversal time for edges. In this scenario the length of the path is a random variable and we consider a formulation with chance constraints, i.e., a bound on the probability that the length of the path exceeds the allotted budget. Our proposed solution casts the problem as an instance of a suitably defined Constrained Markov Decision Process and uses a Lagrangian formulation to solve it. In particular, exploiting some structural properties of the associated decision process we can solve the Markov Decision Process using a Lagrangian approach and efficiently determine the optimal Lagrange multiplier. Our method is experimentally evaluated and demonstrated to be significantly faster than previous solutions using a linear programming approach to solve the Stochastic Orienteering Problem with chance constraints.

The need for comprehensive telemedicine solutions is becoming increasingly relevant due to challenges associated with the ageing population, the increasing shortage of health-care providers, and, more recently, the global pandemic. Existing solutions primarily focus on, e.g., electronic medical records, audiovisual connections, and, in some cases, robotic systems with very basic capabilities. Here we present a fundamentally new, holistic approach to a remote doctor visit, which enables transparent remote examination, anomaly detection, diagnosis, and rehabilitation. Our dual doctor-patient twin paradigm involves two robotic systems: one representing the doctor to the patient ("GARMI") and one representing the patient to the doctor ("MUCKI"). Through bidirectional telepresence control, this system enables transparent, natural, remote haptic interaction between doctor and patient. The control, interaction, and knowledge transfer to the doctor is enhanced by AI-based visual motion and facial expression analysis as well as a digital twin of the patient. Thus, each stage of a doctor visit can be replicated in the context of telemedicine and shared autonomy: from first assessment to observation-based and remote physical examination, to a better-informed doctor diagnosis and robot-assisted telerehabilitation.

Robot grasping and dexterous, in-hand manipulation allow robots to interact with their surroundings and execute a plethora of complex tasks such as pushing buttons, opening doors, and interacting with electrical appliances. In robotics, such complicated tasks are typically executed by multi-fingered end-effectors that are heavy, rigid, and expensive, employing numerous degrees of freedom and actuation. In this paper, we focus on the analysis, design, and development of a multi-grasp, reconfigurable, five fingered, anthropomorphic robot hand that can facilitate the execution of both robust grasping and dexterous manipulation tasks in service robotics and industrial automation applications. The robot hand is composed of eight actuators driving eighteen degrees of freedom with a telescoping mechanism and opposable thumb and pinky fingers to produce multiple anthropomorphic and non-anthropomorphic configurations for grasping and manipulation tasks. The reconfigurable finger base frames allow the hand to transform and utilize its degrees of actuation in an optimal manner to overcome its underactuated limitations. The underactuated robot hand is designed with a human hand structure that takes advantage of objects specifically designed for human operation (e.g., tool or handles with ergonomics for the human hand). This allows the system to better operate within a human-centered environment. The effectiveness of the proposed device is experimentally validated through three different tests: i) grasping experiments involving everyday-life objects, ii) force experiments that assess the force exertion capabilities of the hand in different finger base frame configurations, and iii) demonstration of in-hand object manipulation capabilities. The proposed hand weighs 1.28 kg and has a cost of approximately $1920 USD. The device is capable of exerting up to 14.3 N of contact force during pinch grasping and a maximum of 150.6 N power grasping.

Convolutional Neural Networks (CNNs) have been increasingly applied in visual classification tasks by replacing hand-crafted features with deep features. However, problems such as inter-class similarity and intra-class variation led to the need of obtaining more descriptive features. To accomplish this, a new semantic inter-object relationship approach is proposed, which is based on the distance relationships between recognized objects. This new source of information represents how close or apart objects belonging to two object classes are, which, together with the number of object occurrences, allows to develop a more descriptive semantic feature representation of the scene. To exploit such semantic features, a two-branch CNN architecture based on 1D and 2D convolutional layers, is proposed. Also, an enhancement version, GSF2AppV2, of the Global and Semantic Feature Fusion described in [1] is proposed by integrating the new semantic inter-object relationship approach, as well as the aforementioned two-branch CNN architecture. The GSF2AppV2 is also composed of a CNN-based global feature branch, where five different CNN-based feature extraction approaches were assessed as global feature extraction modules. Moreover, to combine global and semantic features, two feature fusion approaches are proposed and evaluated: correlation and triple concatenation. GSF2AppV2 was evaluated in two benchmark datasets: the SUN RGB-D and NYU Depth V2. State-of-the-art results were achieved on both datasets, showing the effectiveness of the proposed semantic feature approach on the pipeline.

In current object detection, algorithms require the object to be directly visible in order to be detected. As humans, however, we intuitively use visual cues caused by the respective object to already make assumptions about its appearance. In the context of driving, such cues can be shadows during the day and often light reflections at night. In this paper, we study the problem of how to map this intuitive human behavior to computer vision algorithms to detect oncoming vehicles at night just from the light reflections they cause by their headlights. For that, we present an extensive open-source dataset containing 59 746 annotated grayscale images out of 346 different scenes in a rural environment at night. In these images, all oncoming vehicles, their corresponding light objects (e. g., headlamps), and their respective light reflections (e. g., light reflections on guardrails) are labeled. In this context, we discuss the characteristics of the dataset and the challenges in objectively describing visual cues such as light reflections. We provide different metrics for different ways to approach the task and report the results we achieved using state-of-the-art and custom object detection models as a first benchmark. With that, we want to bring attention to a new and so far neglected field in computer vision research, encourage more researchers to tackle the problem, and thereby further close the gap between human performance and computer vision systems.

This article describes an open-source hardware platform for controlling pneumatic soft robotic systems and presents the comparison of control schemes with on-off and proportional valves. The Pneumatic Soft Robotics Driver (PneuSoRD) can be used with up to one pump and pressure accumulator, 26 on-off valves, and 5 proportional valves, any of which can be operated in open or closed-loop control using up to 12 sensor inputs, which allows for the simultaneous control of a large number of soft actuators. The electronic driver connects to a National Instruments myRIO controller or an Arduino Due with the use of an adapter shield. A library of pressure control algorithms in both LabVIEW and Simulink is provided that includes bang-bang control, hysteresis control and PID control using on-off or proportional valves. LabVIEW and Simulink provide user-friendly interfaces for rapid prototyping of control algorithms and real-time evaluation of pressure dynamics. The characteristics and performance of these control methods and pneumatic setups are evaluated to simplify the choice of valves and control algorithm for a given application.