IROS 2021 论文列表

Deep policy networks enable robots to learn behaviors to solve various real-world complex tasks in an end-to-end fashion. However, they lack transparency to provide the reasons of actions. Thus, such a black-box model often results in low reliability and disruptive actions during the deployment of the robot in practice. To enhance its transparency, it is important to explain robot behaviors by considering the extent to which each input feature contributes to determining a given action. In this paper, we present an explicit analysis of deep policy models through input attribution methods to explain how and to what extent each input feature affects the decisions of the robot policy models. To this end, we present two methods for applying input attribution methods to robot policy networks: (1) we measure the importance factor of each joint torque to re ect the influence of the motor torque on the end-effector movement, and (2) we modify a relevance propagation method to handle negative inputs and outputs in deep policy networks properly. To the best of our knowledge, this is the first report to identify the dynamic changes of input attributions of multi-modal sensor inputs in deep policy networks online for robotic manipulation.

Locating odour sources with mobile robots is a difficult task with many real world applications. Over the years, researchers have devised bio-inspired and cognitive methods to enable mobile robots to fulfil this task. One of the most popular cognitive approaches is Infotaxis, which computes a probability map for the location of the chemical source and, on each time step, moves the robot in the direction that minimises the entropy of that probability map. The main difficulty for applying Infotaxis in the real world is selecting proper values for the parameters of its internal gas dispersion model, as it has been shown that its performance is greatly influenced by the accuracy of said model. This work proposes a Genetic Algorithm for optimising those parameters for specific environments. The proposed method is applied to environments with distinct wind and odour dispersion characteristics and the resulting parameters are compared. Moreover, the performance of Infotaxis is compared to that of reactive search strategies evolved by Geometric Syntactic Genetic Programming. The statistically validated results show that the evolved reactive strategies achieve equivalent success rates to Infotaxis, while being significantly faster. Real world experiments conducted in a controlled wind tunnel validated the simulation results.

Event cameras are neuromorphic vision sensors that are able to capture high dynamic range with low latency in microseconds, without motion blur. Their strength lies in the unique representation of data as asynchronous events, enabling detection of scene structures less invariantly from dynamic luminance changes. However, a single event does not represent spatial information, and events must be integrated to translate into meaningful information. Therefore, state-of-the-art deep learning algorithms have focused on reconstructing the original scene from events. However, as environmental variances are also captured throughout events and restored in reconstructed images, simple reconstruction does not help achieving robust visual place recognition. In this paper, we suggest to use reconstructed event edges denoised for place recognition. While brightness wavers with dynamic environmental variances, edge contours only change with gradient magnitude scale. We utilize the high dynamic range of event cameras to detect these scaled edges from different environments and show that using reconstructed edges shows robust performance in overcoming day-to-night illumination variance without a large training set.

Among prerequisites for a synthetic agent to inter-act with dynamic scenes, the ability to identify independently moving objects is specifically important. From an application perspective, nevertheless, standard cameras may deteriorate remarkably under aggressive motion and challenging illumination conditions. In contrast, event-based cameras, as a category of novel biologically inspired sensors, deliver advantages to deal with these challenges. Its rapid response and asynchronous nature enables it to capture visual stimuli at exactly the same rate of the scene dynamics. In this paper, we present a cascaded two-level multi-model fitting method for identifying independently moving objects (i.e., the motion segmentation problem) with a monocular event camera. The first level leverages tracking of event features and solves the feature clustering problem under a progressive multi-model fitting scheme. Initialized with the resulting motion model instances, the second level further addresses the event clustering problem using a spatio-temporal graph-cut method. This combination leads to efficient and accurate event-wise motion segmentation that cannot be achieved by any of them alone. Experiments demonstrate the effectiveness and versatility of our method in real-world scenes with different motion patterns and an unknown number of independently moving objects.

We develop an event-triggered control strategy for a weighted-unbalanced directed homogeneous robot network to reach a dynamic consensus in this work. We present some guarantees for synchronizing a robot network when all robots have access to the reference and when a limited number of robots have access. The proposed event-triggered control can reduce and avoid the periodic updating of the signals. Unlike some current control methods, we prove stability by making use of a logarithmic norm, which extends the possibilities of the control law to be applied to a wide range of directed graphs, in contrast to other works where the event-triggered control can be only implemented over strongly connected and weight-balanced digraphs. We test the performance of our algorithm by carrying out experiments both in simulation and in a real team of robots.

In endoskeleton-type assist suits, since the compressive force is added to the spine, a lumbar disc may be badly affected. However, in general, the lumbar burdens while lifting a heavy object under wearing the endoskeleton-type assist suit have not been fully investigated. In this paper, first, lumbar burdens while lifting the heavy object under wearing the endoskeleton-type assist suit ‘Sustainable’ were investigated through dynamic analysis based on the musculoskeletal model software AnyBody Modeling System. The results of analysis showed that the lumbar burdens can be decreased under wearing the Sustainable as compared with not wearing the Sustainable. Second, in order to improve the utility of the Sustainable, suitable number of artificial muscles of the Sustainable was investigated. Under the condition that the CO2 gas cylinder is used as an air source, Sustainable with two artificial muscles and that with one artificial muscle were compared in terms of assist effect, lumbar burdens and cost performance through measurements of practical assistable times and experiments of lifting the heavy object. Then, suitable number of artificial muscles of the Sustainable was determined.

We compare a state-of-the-art deep image retrieval and a deep place recognition method for place recognition using LiDAR data. Place recognition aims to detect previously visited locations and thus provides an important tool for navigation, mapping, and localisation. Experimental comparisons are conducted using challenging outdoor and indoor datasets, Oxford Radar RobotCar and COLD, in the "long-term" setting where the test conditions differ substantially from the training and gallery data. Based on our results the image retrieval methods using LiDAR depth images can achieve accurate localization (the single best match recall 80%) within 5.00 m in urban outdoors. In office indoors the comparable accuracy is 50 cm but is more sensitive to changes in the environment.

Recent Semantic SLAM methods combine classical geometry-based estimation with deep learning-based object detection or semantic segmentation. In this paper we evaluate the quality of semantic maps generated by state-of-the-art class-and instance-aware dense semantic SLAM algorithms whose codes are publicly available and explore the impacts both semantic segmentation and pose estimation have on the quality of semantic maps. We obtain these results by providing algorithms with ground-truth pose and/or semantic segmentation data available from simulated environments. We establish that semantic segmentation is the largest source of error through our experiments, dropping mAP and OMQ performance by up to 74.3% and 71.3% respectively.

Soft robotic devices, including actuators fabricated from materials with a low modulus of elasticity, such as silicone elastomers, have gained significant interest in recent years. A flexible sensor is a vital component for estimating the conditions of soft actuators, such as shape, and deformation due to contact events. However, it is challenging to develop a flexible sensor with tolerability and versatility for soft actuators. Additionally, when an embedded sensor is employed, the fabrication process becomes complex. Therefore, to have tolerability and to increase versatility, we propose a method for estimating the shape of a soft pneumatic actuator based on vibroacoustic sensing. We employ a data-driven approach by utilizing several machine-learning techniques; hence, the proposed method could be applied to other types of actuators without employing any specific sensor or changing the fabrication process. The convolutional neural network is used as one of the dominant techniques, and huge annotated datasets are required. However, datasets can be obtained automatically using a robotic system, and estimation targets can be changed smoothly by switching datasets. We confirmed the feasibility, and versatility of the proposed method through several evaluation experiments. The bending angle can be estimated in real-time using active vibroacoustic, by emitting sweep signals. The mean error of bending angle estimation was between 5° to 10°. Furthermore, the proposed method is applied to a tensile actuator, and the mean error of the estimated length was approximately 0.21 mm.

Estimating the center of mass (CoM) is essential for both gait planning and controlling of lower limb exoskeletons. Different from CoM estimation in human and humanoid robots, a critical issue in human-exoskeleton systems pis how to describe the effect of physical human-exoskeleton interactions in estimating the CoM of lower limb exoskeletons. This paper presents a novel center of mass estimation method Physically Coupled Serial Chain (PCSC) for human-coupled lower limb exoskeleton systems. Different from traditional serial chain methods, the proposed PCSC involves physical human-exoskeleton models to describe physical interactions between the pilot and the lower limb exoskeleton. We demonstrated the effectiveness of proposed PCSC model in the AIDER lower limb exoskeleton system. Experimental results indicate that the proposed PCSC model is more accuracy than traditional serial chain methods.

Estimating depth from a monocular image is an ill-posed problem: when the camera projects a 3D scene onto a 2D plane, depth information is inherently and permanently lost. Nevertheless, recent work has shown impressive results in estimating 3D structure from 2D images using deep learning. In this paper, we put on an introspective hat and analyze state-of-the-art monocular depth estimation models in indoor scenes to understand these models’ limitations and error patterns. To address errors in depth estimation, we introduce a novel Depth Error Detection Network (DEDN) that spatially identifies erroneous depth predictions in the monocular depth estimation models. By experimenting with multiple state-of-the-art monocular indoor depth estimation models on multiple datasets, we show that our proposed depth error detection network can identify a significant number of errors in the predicted depth maps. Our module is flexible and can be readily plugged into any monocular depth prediction network to help diagnose its results. Additionally, we propose a simple yet effective Depth Error Correction Network (DECN) that iteratively corrects errors based on our initial error diagnosis.

In this study, we propose a predictive model composed of a recurrent neural network including parametric bias and stochastic elements, and an environmentally adaptive robot control method including variance minimization using the model. Robots which have flexible bodies or whose states can only be partially observed are difficult to modelize, and their predictive models often have stochastic behaviors. In addition, the physical state of the robot and the surrounding environment change sequentially, and so the predictive model can change online. Therefore, in this study, we construct a learning-based stochastic predictive model implemented in a neural network embedded with such information from the experience of the robot, and develop a control method for the robot to avoid unstable motion with large variance while adapting to the current environment. This method is verified through a mobile robot in simulation and to the actual robot Fetch.

This paper examines a family of designs for magnetic cubes and counts how many configurations are possible for each design as a function of the number of modules. Magnetic modular cubes are cubes with magnets arranged on their faces. The magnets are positioned so that each face has either magnetic south or north pole outward. Moreover, we require that the net magnetic moment of the cube passes through the center of opposing faces. These magnetic arrangements enable coupling when cube faces with opposite polarity are brought in close proximity and enable moving the cubes by controlling the orientation of a global magnetic field. This paper investigates the 2D and 3D shapes that can be constructed by magnetic modular cubes, and describes all possible magnet arrangements that obey these rules. We select ten magnetic arrangements and assign a "color" to each of them for ease of visualization and reference. We provide a method to enumerate the number of unique polyominoes and polycubes that can be constructed from a given set of colored cubes. We use this method to enumerate all arrangements for up to 20 modules in 2D and 16 modules in 3D. We provide a motion planner for 2D assembly and through simulations compare which arrangements require fewer movements to generate and which arrangements are more common. Hardware demonstrations explore the self-assembly and disassembly of these modules in 2D and 3D.

We propose an energy-efficient controller to minimize the energy consumption of a mobile robot by dynamically manipulating the mechanical and computational actuators of the robot. The mobile robot performs real-time vision-based applications based on an event-based camera. The actuators of the controller are CPU voltage/frequency for the computation part and motor voltage for the mechanical part. We show that independently considering speed control of the robot and voltage/frequency control of the CPU does not necessarily result in an energy-efficient solution. In fact, to obtain the highest efficiency, the computation and mechanical parts should be controlled together in synergy. We propose a fast hill-climbing optimization algorithm to allow the controller to find the best CPU/motor configuration at run-time and whenever the mobile robot is facing a new environment during its travel. Experimental results on a robot with Brushless DC Motors, Jetson TX2 board as the computing unit, and a DAVIS-346 event-based camera show that the proposed control algorithm can save battery energy by an average of 50.5%, 41%, and 30%, in low-complexity, medium-complexity, and high-complexity environments, over baselines.

This paper studies a novel encirclement guaranteed cooperative pursuit problem involving N pursuers and a single evader in an unbounded two-dimensional game domain. Throughout the game, the pursuers are required to maintain encirclement of the evader, i.e., the evader should always stay inside the convex hull generated by all the pursuers, in addition to achieving the classical capture condition. To tackle this challenging cooperative pursuit problem, a robust model predictive control (RMPC) based formulation framework is first introduced, which simultaneously accounts for the encirclement and capture requirements under the assumption that the evader’s action is unavailable to all pursuers. Despite the reformulation, the resulting RMPC problem involves a bilinear constraint due to the encirclement requirement. To further handle such a bilinear constraint, a novel encirclement guaranteed partitioning scheme is devised that simplifies the original bilinear RMPC problem to a number of linear tube MPC (TMPC) problems solvable in a decentralized manner. Simulation experiments demonstrate the effectiveness of the proposed solution framework. Furthermore, comparisons with existing approaches show that the explicit consideration of the encirclement condition significantly improves the chance of successful capture of the evader in various scenarios.

Robotic legs often lag behind the performance of their biological counterparts. The inherent passive dynamics of natural legs largely influences the locomotion and can be abstracted through the spring-loaded inverted pendulum (SLIP) model. This model is often approximated in physical robotic legs using a leg with minimal mass. Our work aims to embed the SLIP dynamics by using a nonlinear strict oscillatory mode into a segmented robotic leg with significant mass, to minimize the control required for achieving periodic motions. For the first time, we provide a realization of a nonlinear oscillatory mode in a robotic leg prototype. This is achieved by decoupling the polar task dynamics and fulfilling the resulting conditions with the physical leg design. Extensive experiments validate that the robotic leg effectively embodies the strict mode. The decoupled leg-length dynamic is exhibited in leg configurations corresponding to the stance and flight phases of the locomotion task, both for the passive system and when actuating the motors. We additionally show that the leg retains this behavior while performing jumping in place experiments.

Exploration and monitoring of hazardous fields in marine environments is one of the most promising tasks to be performed by fleets of low-cost micro autonomous underwater vehicles (μAUVs). In contrast to vehicles in other domains, underwater robots are forced to perform all computations onboard as no powerful communication links are available underwater. This puts the focus on computationally efficient field exploration algorithms. We propose CBTS-GMRF – an extremely light-weight tree-search exploration framework suitable for embedded computing. With our framework we build on recent work in POMDP-exploration and field belief representations based on efficient Gaussian Markov random fields (GMRF). We propose a reward function for energy-efficient field exploration together with a sparse trajectory parameterization. By reducing both, energy consumption and computational complexity, we enable underwater field exploration with μAUVs. We benchmark the performance of our exploration framework in simulation against state-of-the-art exploratory planning schemes and provide an experimental study using a low-cost micro diving agent. In order to support community-wide algorithm benchmarking, our code and robot design can be accessed online.

Most commercially available fixed-wing aerial vehicles (FWV) can carry only small, lightweight computing hardware such as Jetson TX2 onboard. Solving non-linear trajectory optimization on these computing resources is computationally challenging even while considering only the kinematic motion model. Most importantly, the computation time increases sharply as the environment becomes more cluttered. In this paper, we take a step towards overcoming this bottleneck and propose a trajectory optimizer that achieves online performance on both conventional laptops/desktops and Jetson TX2 in a typical urban environment setting. Our optimizer builds on the novel insight that the seemingly non-linear trajectory optimization problem for FWV has an implicit multi-convex structure. Our optimizer exploits these computational structures by bringing together diverse concepts from Alternating Minimization, Bregman iteration, and Alternating Direction Method of Multipliers. We show that our optimizer outperforms the state-of-the-art implementation of sequential quadratic programming approach in optimal control solver ACADO in computation time and solution quality measured in terms of control and goal reaching cost.

Highly dynamic robotic tasks require high-speed and reactive robots. These tasks are particularly challenging due to the physical constraints, hardware limitations, and the high uncertainty of dynamics and sensor measures. To face these issues, it’s crucial to design robotics agents that generate precise and fast trajectories and react immediately to environmental changes. Air hockey is an example of this kind of task. Due to the environment’s characteristics, it is possible to formalize the problem and derive clean mathematical solutions. For these reasons, this environment is perfect for pushing to the limit the performance of currently available general-purpose robotic manipulators. Using two Kuka Iiwa 14, we show how to design a policy for general-purpose robotic manipulators for the air hockey game. We demonstrate that a real robot arm can perform fast-hitting movements and that the two robots can play against each other on a medium-size air hockey table in simulation.