IROS 2020 论文列表

Robot control policies learned in simulation do not often transfer well to the real world. Many existing solutions to this sim-to-real problem, such as the Grounded Action Transformation (GAT) algorithm, seek to correct for- or ground-these differences by matching the simulator to the real world. However, the efficacy of these approaches is limited if they do not explicitly account for stochasticity in the target environment. In this work, we analyze the problems associated with grounding a deterministic simulator in a stochastic real world environment, and we present examples where GAT fails to transfer a good policy due to stochastic transitions in the target domain. In response, we introduce the Stochastic Grounded Action Transformation (SGAT) algorithm, which models this stochasticity when grounding the simulator. We find experimentally-for both simulated and physical target domains-that SGAT can find policies that are robust to stochasticity in the target domain.

Humans use simple probing actions to develop intuition about the physical behavior of common objects. Such intuition is particularly useful for adaptive estimation of favorable manipulation strategies of those objects in novel contexts. For example, observing the effect of tilt on a transparent bottle containing an unknown liquid provides clues on how the liquid might be poured. It is desirable to equip general-purpose robotic systems with this capability because it is inevitable that they will encounter novel objects and scenarios. In this paper, we teach a robot to use a simple, specified probing strategy - stirring with a stick- to reduce spillage when pouring unknown liquids. In the probing step, we continuously observe the effects of a real robot stirring a liquid, while simultaneously tuning the parameters to a model (simulator) until the two outputs are in agreement. We obtain optimal simulation parameters, characterizing the unknown liquid, via a Bayesian Optimizer that minimizes the discrepancy between real and simulated outcomes. Then, we optimize the pouring policy conditioning on the optimal simulation parameters determined via stirring. We show that using stirring as a probing strategy result in reduced spillage for three qualitatively different liquids when executed on a UR10 Robot, compared to probing via pouring. Finally, we provide quantitative insights into the reason for stirring being a suitable calibration task for pouring -a step towards automatic discovery of probing strategies.

This work demonstrates a novel approach to steering a magnetic swimming robot in two dimensions with a single pair of Maxwell coils. By leveraging the curvature of the magnetic field gradient, we achieve motion along two axes. This method allows us to control medical magnetic robots using only existing MRI technology, without requiring additional hardware or posing any additional risk to the patient. We implement a switching time optimization algorithm which generates a schedule of control inputs that direct the swimming robot to a goal location in the workspace. By alternating the direction of the magnetic field gradient produced by the single pair of coils per this schedule, we are able to move the swimmer to desired points in two dimensions. Finally, we demonstrate the feasibility of our approach with an experimental implementation on the millimeter scale and discuss future opportunities to expand this work to the microscale, as well as other control problems and real-world applications.

In this work, we present and explain in detail the design of a new type of pneumatic actuator made of fire hose, the fire hose actuator (FHA), see Fig. 1. We model the force output of this type of actuator and we compare the theoretic results to the data measured on the laboratory test stand.Furthermore, we present the design of a pneumatic rotary drive that is actuated by four of the above-mentioned FHAs. The drive unit features intrinsic compliance and it is capable of high-precision positioning. Due to these characteristics, the design concept of the rotary drive is suitable for the potential use in robotics, especially in human-robot collaboration. In addition, we model the static torque distribution of the rotary drive and we compare the theoretic results to the data measured on the realized laboratory test stand. Moreover, we discuss the most important characteristics of the rotary drive. Hence, we present measurements of the adjustable stiffness and we show that high-precision positioning is possible with the system, reaching in ideal areas every bit of the used 17-bit encoder with a resolution of 0.0027°. Moreover, the drive unit is capable of continuous rotation while the maximum continuous torque possible is found to be 63.1 Nm.

Deep Generative Models such as Generative Adversarial Networks (GAN) and Variational Autoencoders (VAE) have found multiple applications in Robotics, with recent works suggesting the potential use of these methods as a generic solution for the estimation of sampling distributions for motion planning in parameterized sets of environments. In this work we provide a first empirical study of challenges, benefits and drawbacks of utilizing vanilla GANs and VAEs for the approximation of probability distributions arising from sampling-based motion planner path solutions. We present an evaluation on a sequence of simulated 2D configuration spaces of increasing complexity and a 4D planar robot arm scenario and find that vanilla GANs and VAEs both outperform classical statistical estimation by an n-dimensional histogram in our chosen scenarios. We furthermore highlight differences in convergence and noisiness between the trained models and propose and study a benchmark sequence of planar C-space environments parameterized by opened or closed doors. In this setting, we find that the chosen geometrical embedding of the parameters of the family of considered C-spaces is a key performance contributor that relies heavily on human intuition about C-space structure at present. We discuss some of the challenges of parameter selection and convergence for applying this approach with an out-of-the box GAN and VAE model.

Energy sources such as batteries do not decrease in mass after consumption, unlike combustion-based fuels. We present the concept of staging energy sources, i.e. consuming energy in stages and ejecting used stages, to progressively reduce the mass of aerial vehicles in-flight which reduces power consumption, and consequently increases flight time. A flight time vs. energy storage mass analysis is presented to show the endurance benefit of staging to multirotors. We consider two specific problems in discrete staging - optimal order of staging given a certain number of energy sources, and optimal partitioning of a given energy storage mass budget into a given number of stages. We then derive results for a continuously staged case of an internal combustion engine driving propellers. Notably, we show that a multirotor powered by internal combustion has an upper limit on achievable flight time independent of the available fuel mass. Lastly, we validate the analysis with flight experiments on a custom two-stage battery-powered quadcopter. This quadcopter can eject a battery stage after consumption in-flight using a custom-designed mechanism, and continue hovering using the next stage. The experimental flight times match well with those predicted from the analysis for our vehicle. We achieve a 19% increase in flight time using the batteries in two stages as compared to a single stage.

The use of tactile information is one of the most important factors for achieving stable in-grasp manipulation. Especially with low-cost robotic hands that provide low-precision control, robust in-grasp manipulation is challenging. Abundant tactile information could provide the required feed-back to achieve reliable in-grasp manipulation also in such cases. In this research, soft distributed 3-axis skin sensors ("uSkin") and 6-axis F/T (force/torque) sensors were mounted on each fingertip of an Allegro Hand to provide rich tactile information. These sensors yielded 78 measurements for each fingertip (72 measurements from the uSkin and 6 measurements from the 6-axis F/T sensor). However, such high-dimensional tactile information can be difficult to process because of the complex contact states between the grasped object and the fingertips. Therefore, a convolutional neural network (CNN) was employed to process the tactile information. In this paper, we explored the importance of the different sensors for achieving in-grasp manipulation. Successful in-grasp manipulation with untrained daily objects was achieved when both 3-axis uSkin and 6-axis F/T information was provided and when the information was processed using a CNN.

A control algorithm that allows a human model to crawl using a pair of supernumerary robotic limbs (SuperLimbs) is presented. The human model and SuperLimbs are coupled by a compliant harness. This work is inspired by the need for wearable robotic systems that can support workers engaged in fatiguing tasks. The walking policy is developed based on Lyapunov analysis. The volume of the region of attraction (ROA) of the system is used to quantify robustness and identify the optimal harness compliance. Simulation experiments are used to verify the performance of the algorithm. The presented formulation allows us to guarantee stable locomotion under nominal conditions and define robustness against modeling error and perturbations. This study is also the first, that the authors are aware of, to address cooperative crawling between a human and a wearable robotic system with state feedback.

Recently, Differential Dynamic Programming (DDP) and other similar algorithms have become the solvers of choice when performing non-linear Model Predictive Control (nMPC) with modern robotic devices. The reason is that they have a lower computational cost per iteration when compared with off-the-shelf Non-Linear Programming (NLP) solvers, which enables its online operation. However, they cannot handle constraints, and are known to have poor convergence capabilities. In this paper, we propose a method to solve the optimal control problem with control bounds through a squashing function (i.e., a sigmoid, which is bounded by construction). It has been shown that a naive use of squashing functions damage the convergence rate. To tackle this, we first propose to add a quadratic barrier that avoids the difficulty of the plateau produced by the sigmoid. Second, we add an outer loop that adapts both the sigmoid and the barrier; it makes the optimal control problem with the squashing function converge to the original control-bounded problem. To validate our method, we present simulation results for different types of platforms including a multi-rotor, a biped, a quadruped and a humanoid robot.

With its essential role in achieving full autonomy of robot navigation, place recognition has been widely studied with various approaches. Recently, numerous point cloud-based methods with deep learning implementation have been proposed with promising results for their application in outdoor environments. However, their performances are not as promising in indoor spaces because of the high level of occlusion caused by structures and moving objects. In this paper, we propose a point cloud-based place recognition method for crowded indoor spaces. The method consists of voxelizing point clouds in spherical coordinates and defining the occupancy of each voxel in ternary values. We also present SpoxelNet, a neural network architecture that encodes input voxels into global descriptor vectors by extracting the structural features in both fine and coarse scales. It also reinforces its performance in occluded places by concatenating feature vectors from multiple directions. Our method is evaluated in various indoor datasets and outperforms existing methods with a large margin.

We present SplitFusion, a novel dense RGB-D SLAM framework that simultaneously performs tracking and dense reconstruction for both rigid and non-rigid components of the scene. SplitFusion first adopts deep learning based semantic instant segmentation technique to split the scene into rigid or non-rigid surfaces. The split surfaces are independently tracked via rigid or non-rigid ICP and reconstructed through incremental depth map fusion. Experimental results show that the proposed approach can provide not only accurate environment maps but also well-reconstructed non-rigid targets, e.g., the moving humans.

We introduce SplitFlyer-a novel quadcopter with an ability to disassemble into two self-contained bicopters through human assistance. As a subunit, the bicopter is a severely underactuated aerial vehicle equipped with only two propellers. Still, each bicopter is capable of independent flight. To achieve this, we provide an analysis of the system dynamics by relaxing the control over the yaw rotation, allowing the bicopter to maintain its large spinning rate in flight. Taking into account the gyroscopic motion, the dynamics are described and a cascaded control strategy is developed. We constructed a transformable prototype to demonstrate consecutive flights in both configurations. The results verify the proposed control strategy and show the potential of the platform for future research in modular aerial swarm robotics.

Spiking neurons might play a larger role than simply as an efficient signal transmitter. Several studies have demonstrated how movements can be generated using networks of spiking neurons. However, the complexity of spiking neural networks makes their implementation difficult, and the use of spiking neurons in robotics has remained largely impractical. In this paper, we show that the addition of a single layer of spiking neurons can help improve performance on stabilization tasks in dynamically changing environments. In a one-dimensional inverted pendulum stabilization task, the spiking neurons seem to expand the space of usable parameters of the controller. Using a robot arm in 3-D space, the additional layer of spiking neurons suffices to improve performance up to 30% on an inverted pendulum stabilization task. We expect this technique to enhance performance in most stabilization tasks but also tasks that are essentially similar such as reaching tasks and posture control. We also expect the effects of this layer to be greatest when the optimal tuning of control parameters is difficult, such as when the environment is unpredictable and dynamic.

Real-time dense 3D localization and mapping systems are required to enable robotics platforms to interact in and with their environments. Several solutions have used surfel representations to model the world. While they produce impressive results, they require heavy and costly hardware to operate properly. Many of them are also limited to static environments and small inter-frame motions. Whereas most of the state of the art approaches focus on the accuracy of the reconstruction, we assume that many robotics applications do not require a high resolution level in the rebuilt surface and can benefit from a less accurate but less expensive map, so as to gain in run-time and memory efficiency. In this paper we propose a fast RGB-D SLAM articulated around a rough and lightweight 3D representation for dense compact mapping in dynamic indoor environment, targeting mainstream computing platforms. A simple and fast formulation to detect and filter out dynamic elements is also presented. We show the robustness of our system, its low memory requirement and the good performance it enables.

Point-clouds are a popular choice for robotics and computer vision tasks due to their accurate shape description and direct acquisition from range-scanners. This demands the ability to synthesize and reconstruct high-quality point-clouds. Current deep generative models for 3D data generally work on simplified representations (e.g., voxelized objects) and cannot deal with the inherent redundancy and irregularity in point-clouds. A few recent efforts on 3D point-cloud generation offer limited resolution and their complexity grows with the increase in output resolution. In this paper, we develop a principled approach to synthesize 3D point-clouds using a spectral-domain Generative Adversarial Network (GAN). Our spectral representation is highly structured and allows us to disentangle various frequency bands such that the learning task is simplified for a GAN model. As compared to spatial-domain generative approaches, our formulation allows us to generate high-resolution point-clouds with minimal computational overhead. Furthermore, we propose a fully differentiable block to transform from the spectral to the spatial domain and back, thereby allowing us to integrate knowledge from well-established spatial models. We demonstrate that Spectral-GAN performs well for point-cloud generation task. Additionally, it can learn a highly discriminative representation in an unsupervised fashion and can be used to accurately reconstruct 3D objects. Our codes are available at https://github.com/samgregoost/Spectral-GAN/.

Recently, tactile sensing has attracted great interest in robotics, especially for facilitating exploration of unstructured environments and effective manipulation. A detailed understanding of the surface textures via tactile sensing is essential for many of these tasks. Previous works on texture recognition using camera based tactile sensors have been limited to treating all regions in one tactile image or all samples in one tactile sequence equally, which includes much irrelevant or redundant information. In this paper, we propose a novel Spatio-Temporal Attention Model (STAM) for tactile texture recognition, which is the very first of its kind to our best knowledge. The proposed STAM pays attention to both spatial focus of each single tactile texture and the temporal correlation of a tactile sequence. In the experiments to discriminate 100 different fabric textures, the spatially and temporally selective attention has resulted in a significant improvement of the recognition accuracy, by up to 18.8%, compared to the non-attention based models. Specifically, after introducing noisy data that is collected before the contact happens, our proposed STAM can learn the salient features efficiently and the accuracy can increase by 15.23% on average compared with the CNN based baseline approach. The improved tactile texture perception can be applied to facilitate robot tasks like grasping and manipulation.

Recent works have proved that combining spatial and temporal visual cues can significantly improve the performance of various vision-based robotic systems. However, for the ultrasonic sensors used in most robotic tasks (e.g. collision avoidance, localization and navigation), there is a lack of benchmark ultrasonic datasets that consist of spatial and temporal data to verify the usability of spatial and temporal ultrasonic cues. In this paper, we are the first to propose a Spatio-Temporal Ultrasonic Dataset (STUD), which aims to develop the ability of ultrasonic sensors by mining spatial and temporal information from multiple ultrasonic measurements. In particular, we first propose a novel Spatio-Temporal (ST) ultrasonic data gathering scheme, in which an innovatory data instance is designed. Besides, part of the data in the STUD is collected in a robot simulator, in which a well-designed corridor map is utilized to increase the data diversity. Then a selection algorithm is proposed to find a proper length of data sequences to obtain the best description of the navigation environments. Finally, we present an end-to-end learning benchmark model that learns driving policies by extracting spatial and temporal ultrasonic cues from the STUD. With the help of our STUD and this benchmark model, more powerful deep neural networks can be trained for addressing the tasks of indoor navigation or motion planning of mobile robots, which is unachievable by using the existing ultrasonic datasets. Comparison experiments verified the effectiveness of spatial and temporal ultrasonic cues for the robot driving policy learning.

Recent approaches to multi-agent reinforcement learning (MARL) with inter-agent communication have often overlooked important considerations of real-world communication networks, such as limits on bandwidth. In this paper, we propose an approach to learning sparse discrete communication through backpropagation in the context of MARL, in which agents are incentivized to communicate as little as possible while still achieving high reward. Building on top of our prior work on differentiable discrete communication learning, we develop a regularization-inspired message-length penalty term, that encourages agents to send shorter messages and avoid unnecessary communications. To this end, we introduce a variable-length message code that provides agents with a general means of modulating message length while keeping the overall learning objective differentiable. We present simulation results on a partially-observable robot navigation task, where we first show how our approach allows learning of sparse communication behavior while still solving the task. We finally demonstrate our approach can even learn an effective sparse communication behavior from demonstrations of an expert (potentially communication-free) policy.

In semantic mapping, which connects semantic information to an environment map, it is a challenging task for robots to deal with both local and global information of environments. In addition, it is important to estimate semantic information of unobserved areas from already acquired partial observations in a newly visited environment. On the other hand, previous studies on spatial concept formation enabled a robot to relate multiple words to places from bottom-up observations even when the vocabulary was not provided beforehand. However, the robot could not transfer global information related to the room arrangement between semantic maps from other environments. In this paper, we propose SpCoMapGAN, which generates the semantic map in a newly visited environment by training an inference model using previously estimated semantic maps. SpCoMapGAN uses generative adversarial networks (GANs) to transfer semantic information based on room arrangements to a newly visited environment. Our proposed method assigns semantics to the map of an unknown environment using the prior distribution of the map trained in known environments and the multimodal observations made in the unknown environment. We experimentally show in simulation that SpCoMapGAN can use global information for estimating the semantic map and is superior to previous methods. Finally, we also demonstrate in a real environment that SpCoMapGAN can accurately 1) deal with local information, and 2) acquire the semantic information of real places.