IROS 2020 论文列表

The widespread use of fluidic actuation for soft robots creates a high demand for soft pumps and compressors. However, current off-the-shelf pumps are usually rigid, noisy, and cumbersome. As a result, it is hard to integrate most commercial pumps into soft robotic systems, which restricts the autonomy and portability of soft robots. This paper presents the novel design of a soft pump based on bellow structure and super-coiled polymer (SCP) artificial muscles. The pump is flexible, lightweight, modular, scalable, quiet, and low cost. The pumping mechanism and fabrication process of the proposed soft pump is demonstrated. A pump prototype is fabricated to verify the proposed design and characterize its performance. From the characterization results, the pump can reach an output flow rate of up to 54 ml/min and delivers pressure up to 2.63 kPa. The pump has potential applications in untethered soft robots and wearable devices.

We consider the motion planning problem of a hopper navigating a terrain comprising stepping stones while optimizing an energy metric. The most widely used approach of discrete searches (e.g., A-star) cannot handle boundary conditions (e.g., end path constraints on position, velocity). However, continuous optimizations can easily deal with the boundary value problem but are not widely used in motion planning because they are computationally intensive and possibly non-convex when one considers the terrain. Here we use a continuous optimization approach within a model predictive control framework. First, we generate a library comprising initial states at an instant in the locomotion cycle (e.g., apex), the controls (e.g., foot placement, amplitude of force), and the states at the same instant at the next step. Next, we fit these step-to-step models with low order polynomials (typically 2nd or 3rd order). Finally, the planner uses these low order step-to-step models to preview a fixed distance ahead and plans the optimal steps and controls. Thereafter, we implement the plan for the first step, followed by replanning. This process continues until the hopper reaches the end of the terrain. The main contributions are low-order polynomial models for fast computation and incorporation of the complex terrain as a cost function.

In this paper, nonlinear control techniques are exploited to balance an unmanned bicycle with enlarged stability domain. We consider two cases. For the first case when the autonomous bicycle is balanced by the flywheel, the steering angle is set to zero, and the torque of the flywheel is used as the control input. The controller is designed based on the Interconnection and Damping Assignment Passivity Based Control (IDA-PBC) method. For the second case when the bicycle is balanced by the handlebar, the bicycle’s velocity is high, and the flywheel is turned off. The angular velocity of the handlebar is used as the control input and the balance controller is designed based on feedback linearization. In these cases, the global stability of the closed-loop unmanned bicycle is theoretically proved based on Lyapunov theory. The experiments are conducted to validate the efficacy of the proposed nonlinear balance controllers.

In this study, we propose a method of noncontact elastography, which allows us to investigate stiffness of soft structures by combining optical and acoustic modalities. We use optical coherence tomography (OCT) as a means of detecting internal deformation of a sample appearing in response to a mechanical force applied by acoustic radiation pressure. Unlike most of other stiffness sensing, this method can be performed without any contacts between the sample and actuator that generates pressure. To demonstrate the method, we measure the vibration velocity of a uniform phantom made of polyurethane, and characterize the mechanical parameters. We then confirm that the measured and calculated attenuation of the vibration over the depth agree well, which is inaccessible with a conventional laser Doppler vibrometer. This result paves a way to characterize more complex internal structures of soft materials.

This paper describes a calibration method for RGB-D camera networks consisting of not only static overlapping, but also dynamic and non-overlapping cameras. The proposed method consists of two steps: online visual odometry-based calibration and depth image-based calibration refinement. It first estimates the transformations between overlapping cameras using fiducial tags, and bridges non-overlapping camera views through visual odometry that runs on a dynamic monocular camera. Parameters such as poses of the static cameras and tags, as well as dynamic camera trajectory, are estimated in the form of the pose graph-based online landmark SLAM. Then, depth-based ICP and floor constraints are added to the pose graph to compensate for the visual odometry error and refine the calibration result. The proposed method is validated through evaluation in simulated and real environments, and a person tracking experiment is conducted to demonstrate the data integration of static and dynamic cameras.

In this paper, we propose a set-inversion approach to validate the controller of a nonlinear system that should satisfy some state constraints. We introduce the notion of follow set which corresponds to the set of all output vectors such that the desired dynamics can be followed without violating the state-constraints. This follow set can then be used to choose feasible trajectories that a mobile robot will be able to follow. An illustrative example with a robot towing a trailer is presented. This example is motivated by the safe control of a boat towing a marine magnetic sensor to find wrecks.

Performing large step-ups is a challenging task for a humanoid robot. It requires the robot to perform motions at the limit of its reachable workspace while straining to move its body upon the obstacle. This paper presents a non-linear trajectory optimization method for generating step-up motions. We adopt a simplified model of the centroidal dynamics to generate feasible Center of Mass trajectories aimed at reducing the torques required for the step-up motion. The activation and deactivation of contacts at both feet are considered explicitly. The output of the planner is a Center of Mass trajectory plus an optimal duration for each walking phase. These desired values are stabilized by a whole-body controller that determines a set of desired joint torques. We experimentally demonstrate that by using trajectory optimization techniques, the maximum torque required to the full-size humanoid robot Atlas can be reduced up to 20% when performing a step-up motion.

In this paper, we propose a novel contextual bandit algorithm that employs a neural network as a reward estimator and utilizes Shannon entropy regularization to encourage exploration, which is called Shannon entropy regularized neural contextual bandits (SERN). In many learning-based algorithms for robotic grasping, the lack of the real-world data hampers the generalization performance of a model and makes it difficult to apply a trained model to real-world problems. To handle this issue, the proposed method utilizes the benefit of an online learning. The proposed method trains a neural network to predict the success probability of a given grasp pose based on a depth image, which is called a grasp quality. We theoretically show that the SERN has a no regret property. We empirically demonstrate that the SERN outperforms ϵ-greedy in terms of sample efficiency.

Robot navigation in human spaces today largely relies on the construction of precise geometric maps and a global motion plan. In this work, we navigate with only local sensing by using available signage - as designed for humans - in human-made environments such as airports. We propose a formalization of "signage" and define 4 levels of signage that we call complete, fully-specified, consistent and valid. The signage formalization can be used on many space skeletonizations, but we specifically provide an approach for navigation on the medial axis. We prove that we can achieve global completeness guarantees without requiring a global map to plan. We validate with two sets of experiments: (1) with real-world airports and their real signs and (2) real New York City neighborhoods. In (1) we show we can use real-world airport signage to improve on a simple random-walk approach, and we explore augmenting signage to further explore signs' impact on trajectory length. In (2), we navigate in varied sized subsets of New York City to show that, since we only use local sensing, our approach scales linearly with trajectory length rather than freespace area.

In this paper, we propose a novel cluster-based Next-Best-View path planning algorithm to simultaneously explore and inspect large-scale unknown environments with multiple Unmanned Aerial Vehicles (UAVs). In the majority of existing informative path-planning methods, a volumetric criterion is used for the exploration of unknown areas, and the presence of surfaces is only taken into account indirectly. Unfortunately, this approach may lead to inaccurate 3D models, with no guarantee of global surface coverage. To perform accurate 3D reconstructions and minimize runtime, we extend our previous online planner based on TSDF (Truncated Signed Distance Function) mapping, to a fleet of UAVs. Sensor configurations to be visited are directly extracted from the map and assigned greedily to the aerial vehicles, in order to maximize the global utility at the fleet level. The performances of the proposed TSGA (TSP-Greedy Allocation) planner and of a nearest neighbor planner have been compared via realistic numerical experiments in two challenging environments (a power plant and the Statue of Liberty) with up to five quadrotor UAVs equipped with stereo cameras.

In micro aerial vehicle (MAV) operations, the success of a mission is highly dependent on navigation performance, which has raised recent interests on navigation-aware path planning. One of the challenges lies in that optimal motions for successful navigation and the designated mission are often different in unknown, unstructured environments, and only sub-optimality may be obtained in each aspect. We aim to organize a two-MAV team that can effectively execute the mission and simultaneously guarantee navigation quality, which consists of a main-agent responsible for mission and a sub-agent for navigation of the team. Especially, this paper focuses on path planning of the sub-agent to provide navigational assistance to the main-agent using a monocular camera. We adopt a graph-based receding horizon planner to find a dynamically feasible path in order for the sub-agent to help the main-agent's navigation. In this process, we present a metric for evaluating the localization performance utilizing the distribution of the features projected to the image plane. We also design a map management strategy and pose-estimation support mechanism in a monocular camera setup, and validate their effectiveness in two scenarios.

Many areas of scientific interest in planetary exploration, such as lunar pits, icy-moon crevasses, and Martian craters, are inaccessible to current wheeled rovers. Rappelling rovers can safely traverse these steep surfaces, but require techniques to navigate their complex terrain. This dynamic navigation is inherently time-critical and communication constraints (e.g. delays and small communication windows) will require planetary systems to have some autonomy.Autonomous navigation for Martian rovers is well studied on moderately sloped and locally planar surfaces, but these methods do not readily transfer to tethered systems in non-planar 3D environments. Rappelling rovers in these situations have additional challenges, including terrain-tether interaction and its effects on rover stability, path planning and control.This paper presents novel traversability analysis and path planning algorithms for rappelling rovers operating on steep terrains that account for terrain-tether interaction and the unique stability and reachability constraints of a rapelling system. The system is evaluated with a series of simulations and an analogue mission. In simulation, the planner was shown to reliably find safe paths down a 55 degree slope when a stable tether-terrain configuration exists and never recommended an unsafe path when one did not. In a planetary analogue mission, elements of the system were used to autonomously navigate Axel, a JPL rappelling rover, down a 30 degree slope with 95% autonomy by distance travelled over 46 meters.

Models of human behaviour, such as pedestrian flows, are beneficial for safe and efficient operation of mobile robots. We present a new methodology for benchmarking of pedestrian flow models based on the afforded safety of robot navigation in human-populated environments. While previous evaluations of pedestrian flow models focused on their predictive capabilities, we assess their ability to support safe path planning and scheduling. Using real-world datasets gathered continuously over several weeks, we benchmark state-of-the-art pedestrian flow models, including both time-averaged and time-sensitive models. In the evaluation, we use the learned models to plan robot trajectories and then observe the number of times when the robot gets too close to humans, using a predefined social distance threshold. The experiments show that while traditional evaluation criteria based on model fidelity differ only marginally, the introduced criteria vary significantly depending on the model used, providing a natural interpretation of the expected safety of the system. For the time-averaged flow models, the number of encounters increases linearly with the percentage operating time of the robot, as might be reasonably expected. By contrast, for the time-sensitive models, the number of encounters grows sublinearly with the percentage operating time, by planning to avoid congested areas and times.

We present a benchmark for online, video-based depth estimation, a problem that is not covered by the current set of benchmarks for evaluating 3D reconstruction, which focus on offline, batch reconstruction. Online depth estimation from video captured by a moving camera is a key enabling technology for compelling applications in robotics and augmented reality. Inspired by progress in many aspects of robotics due to benchmarks and datasets, we propose a new benchmark called NBVC for evaluating methods for online depth estimation from video. Our benchmark is composed of short video sequences with corresponding high-quality ground truth depth maps, derived from the recent Tanks and Temples dataset. We are hopeful that our work will be instrumental in the development of learning-based algorithms for online depth estimation from video clips, and will also lead to improvements in conventional approaches. In addition to the benchmark, we present a superpixel-based plane sweeping stereo algorithm and use it to investigate various aspects of the problem. The paper contains our initial findings and conclusions.

Learning-based approaches often outperform hand-coded algorithmic solutions for many problems in robotics. However, learning long-horizon tasks on real robot hardware can be intractable, and transferring a learned policy from simulation to reality is still extremely challenging. We present a novel approach to model-free reinforcement learning that can leverage existing sub-optimal solutions as an algorithmic prior during training and deployment. During training, our gated fusion approach enables the prior to guide the initial stages of exploration, increasing sample-efficiency and enabling learning from sparse long-horizon reward signals. Importantly, the policy can learn to improve beyond the performance of the sub-optimal prior since the prior's influence is annealed gradually. During deployment, the policy's uncertainty provides a reliable strategy for transferring a simulation-trained policy to the real world by falling back to the prior controller in uncertain states. We show the efficacy of our Multiplicative Controller Fusion approach on the task of robot navigation and demonstrate safe transfer from simulation to the real world without any fine-tuning. The code for this project is made publicly available at https://sites.google.com/view/mcf-nav/home.

Automated vehicles need to not only perceive their environment, but also predict the possible future behavior of all detected traffic participants in order to safely navigate in complex scenarios and avoid critical situations, ranging from merging on highways to crossing urban intersections. Due to the availability of datasets with large numbers of recorded trajectories of traffic participants, deep learning based approaches can be used to model the behavior of road users. This paper proposes a convolutional network that operates on rasterized actor-centric images which encode the static and dynamic actor-environment. We predict multiple possible future trajectories for each traffic actor, which include position, velocity, acceleration, orientation, yaw rate and position uncertainty estimates. To make better use of the past movement of the actor, we propose to employ temporal convolutional networks (TCNs) and rely on uncertainties estimated from the previous object tracking stage. We evaluate our approach on the public "Argoverse Motion Forecasting" dataset, on which it won the first prize at the Argoverse Motion Forecasting Challenge, as presented on the NeurIPS 2019 workshop on "Machine Learning for Autonomous Driving".

Automation in construction continues to be a topic of interest for many in industry and academia. However, the dynamic environments presented in construction sites prove these tasks to be difficult to automate reliably. This paper proposes a novel method of teleoperation for multiple heterogeneous robots within a construction environment. The system is achieved by creating a virtual reality interface that allows an operator to control multiple robots both synchronously and asynchronously. Feedback is provided from an array of RGBD cameras, force sensors, and precise odometry data. The DRC-Hubo and Spot robot platforms are used for implementation and experimentation. Experiments include useful tasks for construction including item manipulation and item delivery of tools and components. Results demonstrate the feasibility of implementing the system in a construction environment, including trajectory comparisons, task learning curves, and successful multi-robot collaboration.

Leveraging multimodal information with recursive Bayesian filters improves performance and robustness of state estimation, as recursive filters can combine different modalities according to their uncertainties. Prior work has studied how to optimally fuse different sensor modalities with analytical state estimation algorithms. However, deriving the dynamics and measurement models along with their noise profile can be difficult or lead to intractable models. Differentiable filters provide a way to learn these models end-to-end while retaining the algorithmic structure of recursive filters. This can be especially helpful when working with sensor modalities that are high dimensional and have very different characteristics. In contact-rich manipulation, we want to combine visual sensing (which gives us global information) with tactile sensing (which gives us local information). In this paper, we study new differentiable filtering architectures to fuse heterogeneous sensor information. As case studies, we evaluate three tasks: two in planar pushing (simulated and real) and one in manipulating a kinematically constrained door (simulated). In extensive evaluations, we find that differentiable filters that leverage crossmodal sensor information reach comparable accuracies to unstructured LSTM models, while presenting interpretability benefits that may be important for safety-critical systems. We also release an open-source library for creating and training differentiable Bayesian filters in PyTorch, which can be found on our project website: https://sites.google.com/view/multimodalfilter.

Material recognition can help inform robots about how to properly interact with and manipulate real-world objects. In this paper, we present a multimodal sensing technique, leveraging near-infrared spectroscopy and close-range high resolution texture imaging, that enables robots to estimate the materials of household objects. We release a dataset of high resolution texture images and spectral measurements collected from a mobile manipulator that interacted with 144 house-hold objects. We then present a neural network architecture that learns a compact multimodal representation of spectral measurements and texture images. When generalizing material classification to new objects, we show that this multimodal representation enables a robot to recognize materials with greater performance as compared to prior state-of-the-art approaches. Finally, we present how a robot can combine this high resolution local sensing with images from the robot's head-mounted camera to achieve accurate material classification over a scene of objects on a table.