IROS 2020 论文列表

In this paper, we propose an online path planning architecture that extends the model predictive control (MPC) formulation to consider future location uncertainties for safer navigation through cluttered environments. Our algorithm combines an object detection pipeline with a recurrent neural network (RNN) which infers the covariance of state estimates through each step of our MPC's finite time horizon. The RNN model is trained on a dataset that comprises of robot and landmark poses generated from camera images and inertial measurement unit (IMU) readings via a state-of-the-art visualinertial odometry framework. To detect and extract object locations for avoidance, we use a custom-trained convolutional neural network model in conjunction with a feature extractor to retrieve 3D centroid and radii boundaries of nearby obstacles. The robustness of our methods is validated on complex quadruped robot dynamics and can be generally applied to most robotic platforms, demonstrating autonomous behaviors that can plan fast and collision-free paths towards a goal point.

This paper presents a novel risk vector-based near miss prediction and obstacle avoidance method. The proposed method uses the sensor readings about the pose of the other obstacles to infer their motion model (velocity and heading) and, accordingly, adapt the risk assessment and take corrective actions if necessary. Relative vector calculations allow the method to perform in real-time. The algorithm has 1.68 times faster computation performance with less change of motion than other methods and it enables a robot to avoid 25 obstacles in a congested area. Fallback behaviors are also proposed in case of faulty sensors or situation changes. Simulation experiments with parameters inferred from experiments in the ocean with our custom-made robotic boat show the flexibility and adaptability of the proposed method to many obstacles present in the environment. Results highlight more efficient trajectories and comparable safety as other state-of-the-art methods, as well as robustness to failures.

Recent CNN-based optical flow approaches have a separated structure of feature extraction and flow estimation. The core task of optical flow is finding the corresponding points while rich representation is just the key part of such matching problems. However, the prior work usually pays more attention to the design of flow decoder than the feature extraction. In this paper, we present a novel optical flow estimation network to enrich the feature representation of each pyramid level, with a hierarchical dilated architecture and a bottom-up aggregation scheme. In addition, inspired by edge guided classical methods, we bring the edge-aware idea into our approach and propose an edge-aware refinement (EAR) subnetwork to handle motion boundaries. Using the same decoding structure as PWC-Net, our network outperforms it by a large margin and leads all its derivatives both on KITTI-2012 and KITTI-2015. Further performance analysis proves the effectiveness of proposed ideas.

Steerable needles have been widely researched in recent years, and they have multiple potential roles in the medical area. The flexibility and capability of avoiding obstacles allow the steerable needles to be applied in the biopsy, drug delivery and other medical applications that require a high degree of freedom and control accuracy. Radius of Curvature (ROC) of the needle while inserting in the soft tissue is an important parameter for evaluation of the efficacy, and steerability of these flexible needles. For our Fracture-directed Stylet-and-Tube Steerable Needles, it is important to find a relationship among the resultant insertion ROC, pre-set wire shape and the Young’s Modulus of soft tissue to characterize this class of steerable needles. In this paper, an approach is provided for obtaining resultant ROC using stylet and tissue’s mechanical properties. A finite element analysis is also conducted to support the reliability of the model. This work sets the foundation for other researchers to predict the insertion ROC based on the mechanical properties of the needle, and the soft tissue that is being inserted.

There are many design trade-offs when building a magnetic manipulator to control millimeter-scale rotating magnetic swimmers for surgical applications.For example, increasing the magnitude of the flux density generated by the magnetic manipulator increases the torque applied to the swimmer, which could enable performing a wider variety of surgical tasks in the future. However, producing stronger magnetic fields has drawbacks, such as increased active power usage.To produce a quickly rotating field, EMs must be quickly charged and discharged. This results in a low power factor (high reactive power used in comparison with the active power). Adding capacitors in series with the electromagnets improves the power factor because the capacitors can provide reactive power. With this method, larger flux densities can be produced without necessitating an increase of the apparent power delivered by the power supplies.This paper highlights the benefits of using capacitors for the magnetic manipulation of rotating swimmers. Rotating swimmers can be used to remove blood clots. The clot removal rate of resonating magnetic manipulators is measured using a realistic blood clot model. This paper also presents a control method for the currents inside the electromagnets that enable 3D navigation without current sensing.

We propose a centralized control framework to select suitable robots from a heterogeneous pool and place them at appropriate locations to monitor a region for events of interest. In the event of a robot failure, our framework repositions robots in a user-defined local neighborhood of the failed robot to compensate for the coverage loss. If repositioning robots locally fails to attain a user-specified level of desired coverage, the central controller augments the team with additional robots from the pool. The size of the local neighborhood around the failed robot and the desired coverage over the region are two objectives that can be varied to achieve a user-specified balance. We investigate the trade-off between the coverage compensation achieved through local repositioning and the computation required to plan the new robot locations. We also study the relationship between the size of the local neighborhood and the number of additional robots added to the team for a given user-specified level of desired coverage. Through extensive simulations and an experiment with a team of seven quadrotors we verify the effectiveness of our framework. We show that to reach a high level of coverage in a neighborhood with a large robot population, it is more efficient to enlarge the neighborhood size, instead of adding additional robots and repositioning them.

We propose to leverage recent advances in reliable 2D pose estimation with Convolutional Neural Networks (CNN) to estimate the 3D pose of people from depth images in multi-person Human-Robot Interaction (HRI) scenarios. Our method is based on the observation that using the depth information to obtain 3D lifted points from 2D body landmark detections provides a rough estimate of the true 3D human pose, thus requiring only a refinement step. In that line our contributions are threefold. (i) we propose to perform 3D pose estimation from depth images by decoupling 2D pose estimation and 3D pose refinement; (ii) we propose a deep-learning approach that regresses the residual pose between the lifted 3D pose and the true 3D pose; (iii) we show that despite its simplicity, our approach achieves very competitive results both in accuracy and speed on two public datasets and is therefore appealing for multi-person HRI compared to recent state-of-the-art methods.

Understanding spatial relations is a key element for natural human-robot interaction. Especially, a robot must be able to manipulate a given scene according to a human verbal command specifying desired spatial relations between objects. To endow robots with this ability, a suitable representation of spatial relations is necessary, which should be derivable from human demonstrations. We claim that polar coordinates can capture the underlying structure of spatial relations better than Cartesian coordinates and propose a parametric probability distribution defined in polar coordinates to represent spatial relations. We consider static spatial relations such as left of, behind, and near, as well as dynamic ones such as closer to and other side of, and take into account verbal modifiers such as roughly and a lot. We show that adequate distributions can be derived for various combinations of spatial relations and modifiers in a sample-efficient way using Maximum Likelihood Estimation, evaluate the effects of modifiers on the distribution parameters, and demonstrate our representation's usefulness in a pick-and-place task on a real robot.

For robots acting in human-centered environments, the ability to improve based on experience is essential for reliable and adaptive operation; however, particularly in the context of robot failure analysis, experience-based improvement is practically useful only if robots are also able to reason about and explain the decisions they make during execution. In this paper, we describe and analyse a representation of execution-specific knowledge that combines (i) a relational model in the form of qualitative attributes that describe the conditions under which actions can be executed successfully and (ii) a continuous model in the form of a Gaussian process that can be used for generating parameters for action execution, but also for evaluating the expected execution success given a particular action parameterisation. The proposed representation is based on prior, modelled knowledge about actions and is combined with a learning process that is supervised by a teacher. We analyse the benefits of this representation in the context of two actions - grasping handles and pulling an object on a table -such that the experiments demonstrate that the joint relational-continuous model allows a robot to improve its execution based on experience, while reducing the severity of failures experienced during execution.

We present a novel static point cloud map construction algorithm, called Removert, for use within dynamic urban environments. Leaving only static points and excluding dynamic objects is a critical problem in various robust robot missions in changing outdoors, and the procedure commonly contains comparing a query to the noisy map that has dynamic points. In doing so, however, the estimated discrepancies between a query scan and the noisy map tend to possess errors due to imperfect pose estimation, which degrades the static map quality. To tackle the problem, we propose a multiresolution range image-based false prediction reverting algorithm. We first conservatively retain definite static points and iteratively recover more uncertain static points by enlarging the query-to- map association window size, which implicitly compensates the LiDAR motion or registration errors. We validate our method on the KITTI dataset using SemanticKITTI as ground truth, and show our method qualitatively competes or outperforms the human-labeled data (SemanticKITTI) in ambiguous regions.

The simulation of static friction, and especially the effect of stiction, is cumbersome to perform in discrete-time due to its discontinuity at zero velocity and its switching behavior. However, it is essential to achieve reliable simulations of friction to develop compliant torque control algorithms, as they are much disturbed by this phenomenon. This paper takes as a base an elastoplastic model approach for friction, which is free from chattering and drift. It proposes two closed-form solutions that can be used to reliably simulate the effect of stiction consistently with the physics-based Stribeck model. These solutions consider the nonlinearity and velocity dependency, which are main characteristics of lubricated joints. One is directly inspired by the Stribeck nonlinear terms, and the other is a simplified rational approximation. The reliability of this simulation method is shown in simulation, where the consistency and stability are assessed. We also demonstrate the accuracy of these methods by comparing them to experimental data obtained from a robot joint equipped with a high gear reduction harmonic drive.

Asymptotically optimal sampling-based planners require an intelligent exploration strategy to accelerate convergence. After an initial solution is found, a necessary condition for improvement is to generate new samples in the so-called "Informed Set". However, Informed Sampling can be ineffective in focusing search if the chosen heuristic fails to provide a good estimate of the solution cost. This work proposes an algorithm to sample the "Relevant Region" instead, which is a subset of the Informed Set. The Relevant Region utilizes cost-to-come information from the planner's tree structure, reduces dependence on the heuristic, and further focuses the search. Benchmarking tests in uniform and general cost-space settings demonstrate the efficacy of Relevant Region sampling.

This paper proposes a novel method to simultaneously perform relative camera pose estimation and planar reconstruction of a scene from two RGB images. We start by extracting and matching superpixel information from both images and rely on a novel multi-model RANSAC approach to estimate multiple homographies from superpixels and identify matching planes. Ambiguity issues when performing homography decomposition are handled by proposing a voting system to more reliably estimate relative camera pose and plane parameters. A non-linear optimization process is also proposed to perform bundle adjustment that exploits a joint representation of homographies and works both for image pairs and whole sequences of image (vSLAM). As a result, the approach provides a mean to perform a dense 3D plane reconstruction from two RGB images only without relying on RGB-D inputs or strong priors such as Manhattan assumptions, and can be extented to handle sequences of images. Our results compete with keypointbased techniques such as ORB-SLAM while providing a dense representation and are more precise than direct and semi-direct pose estimation techniques used in LSD-SLAM or DPPTAM.

We present a relational graph learning approach for robotic crowd navigation using model-based deep reinforcement learning that plans actions by looking into the future. Our approach reasons about the relations between all agents based on their latent features and uses a Graph Convolutional Network to encode higher-order interactions in each agent’s state representation, which is subsequently leveraged for state prediction and value estimation. The ability to predict human motion allows us to perform multi-step lookahead planning, taking into account the temporal evolution of human crowds. We evaluate our approach against a state-of-the-art baseline for crowd navigation and ablations of our model to demonstrate that navigation with our approach is more efficient, results in fewer collisions, and avoids failure cases involving oscillatory and freezing behaviors.

Energy-efficient mapless navigation is crucial for mobile robots as they explore unknown environments with limited on-board resources. Although the recent deep rein-forcement learning (DRL) approaches have been successfully applied to navigation, their high energy consumption limits their use in several robotic applications. Here, we propose a neuromorphic approach that combines the energy-efficiency of spiking neural networks with the optimality of DRL and benchmark it in learning control policies for mapless navigation. Our hybrid framework, spiking deep deterministic policy gradient (SDDPG), consists of a spiking actor network (SAN) and a deep critic network, where the two networks were trained jointly using gradient descent. The co-learning enabled synergistic information exchange between the two networks, allowing them to overcome each other's limitations through a shared representation learning. To evaluate our approach, we deployed the trained SAN on Intel's Loihi neuromorphic processor. When validated on simulated and real-world complex environments, our method on Loihi consumed 75 times less energy per inference as compared to DDPG on Jetson TX2, and also exhibited a higher rate of successful navigation to the goal, which ranged from 1% to 4.2% and depended on the forward-propagation timestep size. These results reinforce our ongoing efforts to design brain-inspired algorithms for controlling autonomous robots with neuromorphic hardware.

Path following is a challenging task for legged robots. In this paper, we present a hierarchical control architecture for path following of a quadruped salamander-like robot, in which, the tracking problem is decomposed into two sub-tasks: high-level policy learning based on the framework of reinforcement learning (RL) and low-level traditional controller design. More specifically, the high-level policy is learned in a physics simulator with a low-level controller designed in advance. To improve the tracking accuracy and to eliminate static errors, a soft Actor-Critic algorithm with state integral compensation is proposed. Additionally, to enhance the generalization and transferability, a compact state representation, which only contains the information of the target path and the abstract action similar to front-back and left-right, is proposed. The proposed algorithm is trained offline in the simulation environment and tested on the self-developed real quadruped salamander-like robot for different path following tasks. Simulation and experiments results validate the satisfactory performance of the proposed method.

One problem in real-world applications of reinforcement learning is the high dimensionality of the action search spaces, which comes from the combination of actions over time. To reduce the dimensionality of action sequence search spaces, macro actions have been studied, which are sequences of primitive actions to solve tasks. However, previous studies relied on humans to define macro actions or assumed macro actions to be repetitions of the same primitive actions. We propose encoded action sequence reinforcement learning (EASRL), a reinforcement learning method that learns flexible sequences of actions in a latent space for a high-dimensional action sequence search space. With EASRL, encoder and decoder networks are trained with demonstration data by using variational autoencoders for mapping macro actions into the latent space. Then, we learn a policy network in the latent space, which is a distribution over encoded macro actions given a state. By learning in the latent space, we can reduce the dimensionality of the action sequence search space and handle various patterns of action sequences. We experimentally demonstrate that the proposed method outperforms other reinforcement learning methods on tasks that require an extensive amount of search.

Robots can learn to do complex tasks in simulation, but often, learned behaviors fail to transfer well to the real world due to simulator imperfections (the "reality gap"). Some existing solutions to this sim-to-real problem, such as Grounded Action Transformation (gat), use a small amount of real-world experience to minimize the reality gap by "grounding" the simulator. While very effective in certain scenarios, gat is not robust on problems that use complex function approximation techniques to model a policy. In this paper, we introduce Reinforced Grounded Action Transformation (rgat), a new sim-to-real technique that uses Reinforcement Learning (RL) not only to update the target policy in simulation, but also to perform the grounding step itself. This novel formulation allows for end-to-end training during the grounding step, which, compared to gat, produces a better grounded simulator. Moreover, we show experimentally in several MuJoCo domains that our approach leads to successful transfer for policies modeled using neural networks.

This paper presents an experimental study to investigate how humans interact with a robotic arm simulating primarily unstable, damping-defined, mechanical environments, and to quantify lower bounds of robotic damping that humans can stably interact with. Human subjects performed posture maintenance tasks while a robotic arm simulated a range of negative damping-defined environments and transiently perturbed the human arm to challenge postural stability. Analysis of 2-dimensional kinematic responses in both the time domain and phase space allowed us to evaluate stability of the coupled human-robot system in both anterior-posterior (AP) and medial-lateral (ML) directions, and to determine the lower bounds of robotic damping for stable physical human-robot interaction (pHRI). All subjects demonstrated higher capacity to stabilize their arm against negative damping-defined environments in the AP direction than the ML direction, evidenced by all 3 stability measures used in this study. Further, the lower bound of robotic damping for stable pHRI was more than 3.5 times lower in the AP direction than the ML direction: -30.0 Ns/m and -8.2 Ns/m in the AP and ML directions, respectively. Sensitivity analysis confirmed that the results in this study were relatively insensitive to varying experimental conditions. Outcomes of this study would allow us to design a less conservative robotic impedance controller that utilizes a wide range of robotic damping, including negative damping, and achieves more transparent and agile operations without compromising coupled stability and safety of the human-robot system, and thus improves the overall performance of pHRI.