IROS 2017 论文列表

Teleoperating a quadrotor equipped with a monocular camera for exploring a wide area in search of something has become a common practice in many application scenarios (e.g. search and rescue). In order to efficiently plan operations, estimating the 3D pose of a point of interest is as important as detecting it. In this paper we propose a novel bilateral teleoperation architecture where an estimation scheme is exploited for recovering the position of a set of visual features while an operator steers the motion of the quadrotor UAV. The operator acts on a force-feedback master device that produces force cues meant to suggest where to drive the quadrotor for improving the convergence rate of the estimation process. The effectiveness of the proposed teleoperation strategy is validated by means of hardware in the loop simulations.

This paper addresses robot task model learning and planning with loops. By detecting and modeling loops in solved tasks it is possible to learn and solve wider classes of problems. We extend our previous work on experience-based planning domains in robotics to detect, represent and generate loops in action sequences. This approach provides methods for, (i) conceptualizing robot experiences possibly containing loops and learning high-level robot activity schemata with loops; and (ii) instantiating schemata with loops for solving problem instances of the same task with varying sets of objects. Demonstrations of this system in both real and simulated environments prove its potentialities.

Autonomous Science is a field of study which aims to extend the autonomy of exploration robots from low level functionality, such as on-board perception and obstacle avoidance, to science autonomy, which allows scientists to specify missions at task level. This will enable more remote and extreme environments such as deep ocean and other planets to be studied, leading to significant science discoveries. This paper presents an approach to extend the high level autonomy of robots by enabling them to model and reason about scientific knowledge on-board. We achieve this by using Bayesian networks to encode scientific knowledge and adapting Monte Carlo Tree Search techniques to reason about the network and plan informative sensing actions. The resulting knowledge representation and reasoning framework is anytime, handles large state spaces and robust to uncertainty making it highly applicable to field robotics. We apply the approach to a Mars exploration mission in which the robot is required to plan paths and decide when to use its sensing modalities to study a scientific latent variable of interest. Extensive simulation results show that our approach has significant performance benefits over alternative methods. We also demonstrate the practicality of our approach in an analog Martian environment where our experimental rover, Continuum, plans and executes a science mission autonomously.

This paper presents an anticipative robot kinematic limitation avoidance algorithm for collaborative robots. The main objective is to improve the performance and the intuitivity of physical human-robot interaction. Currently, in such interactions, the human user must focus on the task as well as on the robot configuration. Indeed, the user must pay a close attention to the robot in order to avoid limitations such as joint position limitations, singularities and collisions with the environment. The proposed anticipative algorithm aims at relieving the human user from having to deal with such limitations by automatically avoiding them while considering the user's intentions. The framework developed to manage several limitations occurring simultaneously in three-dimensional space is first presented. The algorithm is then presented and detailed for each individual limitation of a spatial RRR serial robot. Finally, experiments are performed in order to assess the performance of the algorithm.

Unmanned Aerial Vehicle (UAV) are increasingly used as data collectors for Wireless Sensors Networks (WSN) on the ground. Most of current research proposes optimizations for itinerary creation for a single UAV. Contrary to this, our work proposes a distributed algorithm for collecting WSN data using a dynamic set of UAVs, that takes into account that UAVs leave or join the group due to recharging or malfunctions. In our work, we consider that the UAVs only have medium-range communication capability (a few meters) to deliver collected data, similar to the assumptions in related work. Compared to the costly and non-real-time Traveling Salesman Problem (TSP)-approach our algorithm delivers approximately 3% more efficient sensor visiting on certain scenarios without using the optimized tour.

A robust state estimator is presented by fusing the aerodynamic model of multi-rotor UAVs with measurements from optical flow and other low-cost sensors such as IMU, magnetometer, and ultrasonic sensor. Due to the particular aerodynamics of multi-rotor UAVs, the body velocity in the rotor plane is able to be measured by the accelerometer. We therefore propose a novel state estimator by fully exploring the characteristic of aerodynamics of multi-rotor UAV. Our state estimator is fast and easy to be implemented. We have tested our estimator with different platforms in different scenes. Experimental results show that our estimator performs robustly in low light conditions where existing methods usually fail.

This paper presents a method for navigating 3D dynamically walking bipedal robots amidst obstacles. Our framework relies on composing gait primitives corresponding to limit-cycle locomotion behaviors and it produces nominal motion plans that are compatible with the system's dynamics and can be tracked with high fidelity. The low-level controllers of the biped are designed within the Hybrid Zero Dynamics (HZD) framework. Exploiting the dimensional reduction afforded by HZD and properties of invariant sets of switching systems among multiple equilibria, we obtain polynomial approximations of a reduced order Poincaré map and of the net change of the center of mass location over a stride. These polynomials are then incorporated in a high-level Rapidly Exploring Random Tree (RRT) planner to generate nominal plans which are tracked by the biped with drastically low drifting errors, without adversely affecting the time for computation.

This paper reports on a novel two-step algorithm for the estimation of full 6-degree-of-freedom (DOF) [tx, ty, tz, θx, θy, θz] rigid body transformation between any two overlapping point-clouds that have a dominant ground plane. We first estimate the ground plane (X-Y plane) from the two 3D point-clouds and align them to obtain a good estimate of the distance between the ground planes (i.e. tz) and rotations θx and θy about the X and Y axis respectively using the Rodrigues rotation formula. The remaining parameters (tx, ty, θz) are then estimated by maximizing the total mutual information (MI) between the 2D feature maps generated from the multi-modal sensor data. Experimental results using scans obtained by a vehicle equipped with a 3D laser scanner and an omnidirectional camera are used to validate the robustness of the proposed algorithm over a wide range of initial conditions. The proposed method provides an efficient framework for multi-modal sensor data fusion and provides a robust solution to the scan alignment problem.

There are driving applications for large populations of tiny robots in robotics, biology, and chemistry. These robots often lack onboard computation, actuation, and communication. Instead, these “robots” are particles carrying some payload and the particle swarm is controlled by a shared control input such as a uniform magnetic gradient or electric field. In previous works, we showed that the 2D position of each particle in such a swarm is controllable if the workspace contains a single obstacle the size of one particle. Requiring a small, rigid obstacle suspended in the middle of the workspace is a strong constraint, especially in 3D. This paper relaxes that constraint, and provides position control algorithms that only require non-slip wall contact in 2D. Both in vivo and artificial environments often have such boundaries. We assume that particles in contact with the boundaries have zero velocity if the shared control input pushes the particle into the wall. This paper provides a shortest-path algorithm for positioning a two-particle swarm, and a generalization to positioning an n-particle swarm. Results are validated with simulations and a hardware demonstration.

This paper presents an algebraic approach for the estimation and control of single-link flexible joint robots. Typically, a sensory set to directly measure all robot states is hardly available. Especially in cost-efficient robotic systems as widely used in the industry, robots do not profit from ‘link-sided’ position measurements for end-effector control tasks but instead rely on easily accessible ‘motor-sided’ measurements in combination with model-based control strategies. In this paper, we are interested in obtaining online estimates of the unmeasured link position and speed for use in control. To that end, we introduce an algebraic estimation approach which relies only on motor position sensing and an accelerometer mounted near the end-effector. The estimation is based on an algebraic reformulation of the equations of motion which are linear in the unknown variables allowing an estimation process which is of non-asymptotic fashion. The proposed solution is illustrated by simulation and experimentally with an industrial robot. The robot is controlled by a common 2-DOF structure consisting of feed-forward and feedback control loops. Our experimental results show good tracking performance, especially in terms of suppressing residual vibrations.

Quadrotors that navigate through unknown, cluttered environments have only recently begun to emerge following the development of small form-factor sensing and computing hardware and computationally efficient collision avoidance algorithms. Computation time of planning algorithms has significantly decreased in part to using local information as opposed to using a global map for collision avoidance. Safe planning with local information, however, restricts the direction of travel to remain within the perception system's field-of-view (FOV). The vehicle's motion becomes more constrained with body-mounted narrow FOV sensors, reducing vehicle maneuverability and speed. This work presents a relaxed-constraint Model Predictive Control framework that allows motions outside the perception FOV with guaranteed safety. The key aspect of this approach is the ability to safely choose a motion primitive generated in the past. Simulation and hardware results shows the new framework improves time to goal and flight path efficiency in environments with varying levels of clutter.

In this paper, we propose a novel Aerial Social Force Model (ASFM) that allows autonomous flying robots to accompany humans in urban environments in a safe and comfortable manner. To date, we are not aware of other state-of-the-art method that accomplish this task. The proposed approach is a 3D version of the Social Force Model (SFM) for the field of aerial robots which includes an interactive human-robot navigation scheme capable of predicting human motions and intentions so as to safely accompany them to their final destination. ASFM also introduces a new metric to fine-tune the parameters of the force model, and to evaluate the performance of the aerial robot companion based on comfort and distance between the robot and humans. The presented approach is extensively validated in diverse simulations and real experiments, and compared against other similar works in the literature. ASFM attains remarkable results and shows that it is a valuable framework for social robotics applications, such as guiding people or human-robot interaction.

Policy search methods in reinforcement learning have demonstrated success in scaling up to larger problems beyond toy examples. However, deploying these methods on real robots remains challenging due to the large sample complexity required during learning and their vulnerability to malicious intervention. We introduce Adversarially Robust Policy Learning (ARPL), an algorithm that leverages active computation of physically-plausible adversarial examples during training to enable robust policy learning in the source domain and robust performance under both random and adversarial input perturbations. We evaluate ARPL on four continuous control tasks and show superior resilience to changes in physical environment dynamics parameters and environment state as compared to state-of-the-art robust policy learning methods. Code, data, and additional experimental results are available at: stanfordvl.github.io/ARPL.

Impedance control is commonly implemented for robotic contact applications. Its performance is a function of the dynamic coupling between the environment and the robot's impedance controller: inertia, stiffness and damping. An interaction task may be considered successfully accomplished when the elected performance criteria, such as rise time, overshoot and accommodation are achieved, all in addition to guaranteeing stability. The selection and online changes of impedance parameters for such purpose is considered challenging for practical applications, given the non modeled dynamics of the robot and its intrinsic impedance, uncertainties in the dynamics of environment and tool. This study proposes the usage of a multi-objective genetic algorithm to obtain enhanced impedance controller gains, with the purpose of providing impedance control adaptability to a robot in a real physical interaction application. In this context, a one-degree-of-freedom mechanical contact model is proposed to obtain the values of the objective function. A genetic algorithm is used to obtain the best gains, as this is a nonlinear task. Simulation and experimental results are presented and discussed validating the system performance, which has shown a convergence of less than 6 generations. Tendencies to convergence of controller parameters were observed and discussed.

Appearance changes due to weather and seasonal conditions represent a strong impediment to the robust implementation of machine learning systems in outdoor robotics. While supervised learning optimises a model for the training domain, it will deliver degraded performance in application domains that underlie distributional shifts caused by these changes. Traditionally, this problem has been addressed via the collection of labelled data in multiple domains or by imposing priors on the type of shift between both domains. We frame the problem in the context of unsupervised domain adaptation and develop a framework for applying adversarial techniques to adapt popular, state-of-the-art network architectures with the additional objective to align features across domains. Moreover, as adversarial training is notoriously unstable, we first perform an extensive ablation study, adapting many techniques known to stabilise generative adversarial networks, and evaluate on a surrogate classification task with the same appearance change. The distilled insights are applied to the problem of free-space segmentation for motion planning in autonomous driving.

In this paper, we introduce a walking load control strategy with a cane-type robot for gait training. The population of elderly people is increasing in most of the developed countries, and there is a growing demand for caregivers. Assistive robots are expected as a solution to provide aid in order to reduce the burden of caregivers in gait training of elderly people. Our research group has been developing a series of cane-type walking assistive robots named Intelligent Cane on admittance control to provide safe and efficient gait training. We propose an adaptive design method of the admittance control model in order to control user's walking load for the purpose of rapid improvement of user's physical strength. In the proposed method, the cane robot automatically adjusts control parameters of an admittance control model based on a correlation between a set of admittance parameters and a user's walking energy consumption. We conducted experiments to evaluate the proposed adaptive control strategy through a convergence error between estimated walking load and target one while walking training. We confirm the advantage of using the cane robot in rehabilitation.

An adaptive trajectory tracking problem for a spherical rolling robot driven by a 2DOF pendulum is considered in this paper. A feedback controller is proposed for the goal of tracking the trajectory for the full configuration of the spherical robot. To deal with time-varying uncertainty of the system dynamics, an adaptation mechanism is included in the trajectory tracking controller by parameterizing the system uncertainty with a polynomial function, weighted by unknown constant parameters adjustable by the defined update law. The constructed controller is first tested for the planar hoop-pendulum and then applied to the ball-pendulum system. The convergence to the desired trajectories under the control law is proved and then verified by simulations for tracking circular motions under perturbation.

In this work, we present an adaptive perception method to improve the performance in accuracy and speed of a tactile exploration task. This work extends our previous studies on sensorimotor control strategies for active tactile perception in robotics. First, we present the active Bayesian perception method to actively reposition a robot to accumulate evidence from better locations to reduce uncertainty. Second, we describe the adaptive perception method that, based on a forward model and a predicted information gain approach, allows to the robot to analyse `what would have happened' if a different decision `would have been made' at previous decision time. This approach permits to adapt the active Bayesian perception process to improve the performance in accuracy and reaction time of an exploration task. Our methods are validated with a contour following exploratory procedure with a touch sensor. The results show that the adaptive perception method allows the robot to make sensory predictions and autonomously adapt, improving the performance of the exploration task.

Modular self-reconfigurable robots (MSRs) are mostly used in environments where it is difficult to navigate and explore otherwise. Especially, the shape-changing ability of MSRs makes them more dexterous in these situations compared to fixed-body robots. But when the MSR forms a new configuration, usually, the locomotion pattern for that particular configuration is not known by its constituting robotic modules. The main challenge for modules is to learn how to move in that specific configuration within a reasonable amount of time. In this paper, we study the problem where an MSR needs to learn its movement pattern on-the-fly. To solve this problem, we have proposed a game theoretic solution based on multi-agent reinforcement learning using which the constituting modules distributedly learn the best actions that they need to perform to travel more distance in less time. We have implemented this approach in simulation on both ModRED and Yamor MSR platforms. Results show that our approach performs better (up to 7.86 times) in terms of average speed achieved for most of the tested configurations as compared to an existing locomotion learning approach.