IROS 2019 论文列表

We present an approach to enhance wheeled planetary rover dead-reckoning localization performance by leveraging the use of zero-type constraint equations in the navigation filter. Without external aiding, inertial navigation solutions inherently exhibit cubic error growth. Furthermore, for planetary rovers that are traversing diverse types of terrain, wheel odometry is often unreliable for use in localization, due to wheel slippage. For current Mars rovers, computer vision-based approaches are generally used whenever there is a high possibility of positioning error; however, these strategies require additional computational power, energy resources, adequate features in the environment, and significantly slow down the rover traverse speed. To this end, we propose a navigation approach that compensates for the high likelihood of odometry errors by providing a reliable navigation solution that leverages non-holonomic vehicle constraints as well as state-aware pseudo-measurements (e.g., zero velocity and zero angular rate) updates during periodic stops. By using this, computationally expensive visual-based corrections could be performed less often. Experimental tests that compare against GPS-based localization are used to demonstrate the accuracy of the proposed approach. The source code, post-processing scripts, and example datasets associated with the paper are published in a public repository.

Cable-driven segmented manipulators (CDSMs) featured by superior dexterity, light and slender body are excellent candidates for operations in confined environments. However, the stiffness and load capacity of such manipulators have been a challenge due to their structural elasticity. In this paper, we propose an improved mechanism design based on the preliminary work to enhance the linkage accuracy and arm continuity without sacrificing the dexterity, high stiffness and load capacity of CDSM. The manipulator is composed of 4 improved hybrid active-passive linkage segments. Its short and long linkage cables with pretension mechanism are designed to keep equal angles of adjacent joints. An improved separable small driving control box is also designed with both quick release and load mechanism and stroke amplification mechanism. Then the size of control box can still remain small, even the number of segments and the joint limit angles increase. Considering the improved in-segment linkage characteristic, traditional kinematic equations and Jacobian matrix are greatly simplified with Denavit-Hartenberg (D-H) method. Further trajectory tracking planning based on the simplified kinematics solved the Cartesian space planning for task design. Finally, a prototype system is developed to perform the linkage accuracy and comprehensive obstacle avoidance experiments. Experimental results show that the developed hybrid active and passive CDSM has relatively high accuracy and super dexterity.

Learning from Demonstration (LfD) has been identified as an effective method for making robots adapt to a similar kind of tasks. In this work, a framework of learning from demonstration has been proposed for modelling robot motions. We present an approach based on dimension ascending to learn a dynamical system, so that the reproduced motions can closely follow the demonstrations. In addition, the reproductions can ultimately reach and stop at the target, which reflects the robustness of the method. Therefore, the system accuracy and stability can be better guaranteed simultaneously. The effectiveness of the proposed approach is verified by performing handwriting experiments on the LASA data set.

Model-free reinforcement learning algorithms such as Deep Deterministic Policy Gradient (DDPG) often require additional exploration strategies, especially if the actor is of deterministic nature. This work evaluates the use of model-based trajectory optimization methods used for exploration in Deep Deterministic Policy Gradient when trained on a latent image embedding. In addition, an extension of DDPG is derived using a value function as critic, making use of a learned deep dynamics model to compute the policy gradient. This approach leads to a symbiotic relationship between the deep reinforcement learning algorithm and the latent trajectory optimizer. The trajectory optimizer benefits from the critic learned by the RL algorithm and the latter from the enhanced exploration generated by the planner. The developed methods are evaluated on two continuous control tasks, one in simulation and one in the real world. In particular, a Baxter robot is trained to perform an insertion task, while only receiving sparse rewards and images as observations from the environment.

This work presents a successful implementation of a nonlinear optimization-based Regularized Predictive Control (RPC) for legged locomotion on the MIT Cheetah 3 robot platform. Footstep placements and ground reaction forces at the contact feet are simultaneously solved for over a prediction horizon in real-time. Often in academic literature not enough attention is given to the implementation details that make the theory work in practice and many times it is precisely these details that end up being critical to the success or failure of the theory in real world applications. Nonlinear optimization for real-time legged locomotion control in particular is one of the techniques that has shown promise, but falls short when implemented on hardware systems subjected to computation limits and undesirable local minima. We discuss various algorithms and techniques developed to overcome some of the challenges faced when implementing nonlinear optimization-based controllers for dynamic legged locomotion.

Natural dynamic controllers aim to perform the desired task by exploiting the natural dynamics of the system. This can be accomplished by generating torque patterns to actuate the system rather than accurately following a predefined trajectory. We have previously demonstrated natural dynamic control of compass-biped in simulations. Here we demonstrate successful implementation of this dynamic controller on an under-actuated compass-biped robot. The parameters of the controller, and in particular the magnitude and timing of torque primitives, were optimized using multi-objective optimization via genetic algorithm, accounting for speed and energy efficiency. While the current implementation is in open-loop, this strategy can be extended to include feedback to enhance walking over a wide range of terrains. This proof-of concept provides the basis for future extensions to more complex robots.

Central-place foraging (CPF) is a canonical task in collective robotics with applications to planetary exploration, automated mining, warehousing, and search and rescue operations. We compare the performance of three Central-Place Foraging Algorithms (CPFAs), variants of which have been shown to work well in real robots: spiral-based, rotating-spoke, and random-ballistic. To understand the difference in performance between these CPFAs, we define the price of ignorance and show how this metric explains our previously published empirical results. We obtain upper-bounds for expected complete collection times for each algorithm and evaluate their performance in simulation. We show that site-fidelity (i.e. returning to the location of the last found target) and avoiding search redundancy are key-factors that determine the efficiency of CPFAs. Our formal analysis suggests the following efficiency ranking from best to worst: spiral, spoke, and the stochastic ballistic algorithm.

When disagreements arise in hierarchical relationships, relationship members sometimes prefer conflict management strategies that avoid or quickly end the overt conflict even if the relationship is left in a state of dissatisfaction. Our lab has proposed that a peripheral robotic agent may be able to support these types of relationships during conflict. In this paper, we present the results of an IRB-approved human-robot interaction study that examines how the members of a hierarchical relationship involved in conflict respond to the presence of an unengaged robot. This study serves as a baseline for additional studies. The unengaged robot appears to have a minimal influence on the interaction. The observed conflicts followed the patterns typically described in mediation literature. Our lab previously proposed a computational model to identify weakness and alienation in these relationships. We discuss a partial implementation of this model, and its ability to recognize problems in certain relationships within the data collected. Based on our observations, and the performance of the model’s partial implementation, we suggest considerations that need to be made for an intervening robotic agent.

Social interaction between a robot and rats is important since the robot can generate reproducible social behaviors across trials. However, lacking internal state feedback from the rat makes current robot-rat interaction a very preliminary level comparing with rat-rat interaction. Previous biological studies showed that ultrasonic vocalizations (USVs) emitted by a rat are expressions of its internal emotional states, which therefore can be used as part of feedback for a robot-rat interaction. The challenge is to accurately identify rat USVs in real-time from mix sounds generated by the robot in a noisy environment. To address these problems, we propose an SVM-based rat USVs identification method. This SVM method uses three types of features to represents the characteristics of mix sound and use these multidimensional features to identify rat USVs. Results show that our identification method has an accuracy of 84.29% with only 4.84% false-positive rate. Furthermore, we carefully design the filter window length with respect to sound chunk length and use only one microphone to record the mix sound. All of these efforts are to reduce the calculation time to realize real-time identification. Eventually, the identification process can be executed within 3. 5ms, which definitely meet the real-time demand. This research lays the foundation of the feedback based interaction between rat and robot, and also shows promise in the study of ethology and the interaction between robot and animals.

Vision, as an inexpensive yet information rich sensor, is commonly used for perception on autonomous mobile robots. Unfortunately, accurate vision-based perception requires a number of assumptions about the environment to hold - some examples of such assumptions, depending on the perception algorithm at hand, include purely lambertian surfaces, texture-rich scenes, absence of aliasing features, and refractive surfaces. In this paper, we present an approach for introspective vision for obstacle avoidance (IVOA) - by leveraging a supervisory sensor that is occasionally available, we detect failures of stereo vision-based perception from divergence in plans generated by vision and the supervisory sensor. By projecting the 3D coordinates where the plans agree and disagree onto the images used for vision-based perception, IVOA generates a training set of reliable and unreliable image patches for perception. We then use this training dataset to learn a model of which image patches are likely to cause failures of the vision-based perception algorithm. Using this model, IVOA is then able to predict whether the relevant image patches in the observed images are likely to cause failures due to vision (both false positives and false negatives). We empirically demonstrate with extensive real-world data from both indoor and outdoor environments, the ability of IVOA to accurately predict the failures of two distinct vision algorithms.

In urban driving scenarios, forecasting future trajectories of surrounding vehicles is of paramount importance. While several approaches for the problem have been proposed, the best-performing ones tend to require extremely detailed input representations (e.g. image sequences). As a result, such methods do not generalize to datasets they have not been trained on. In this paper, we propose intermediate representations that are particularly well-suited for future prediction. As opposed to using texture (color) information from images, we condition on semantics and train an autoregressive model to accurately predict future trajectories of traffic participants (vehicles) (see Fig. above). We demonstrate that semantics provide a significant boost over techniques that operate over raw pixel intensities/disparities. Uncharacteristic of state-of-the-art approaches, our representations and models generalize across different sensing modalities (stereo imagery, LiDAR, a combination of both), and also across completely different datasets, collected across several cities, and also across countries where people drive on opposite sides of the road (left-handed vs right-handed driving). Additionally, we demonstrate an application of our approach in multi-object tracking (data association). To foster further research in transferable representations and ensure reproducibility, we release all our code and data.33More qualitative and quantitative results, along with code and data can be found at https://rebrand.ly/INFER-results.

The use of the tendon-sheath mechanism (TSM) is common in flexible surgery robots, because it can flexibly work in limited spaces and provides efficient power transmission. However, hysteresis from nonlinearities such as friction and backlash poses a challenge in controlling precise motion in the surgical instrument. Moreover, this hysteresis is also affected by changes in the various configurations of sheath which limits traditional model-based compensation approaches. Recently, feedback approach using an endoscopic camera is presented, but they use markers which are not appropriate for applying to a real surgical instruments. In this paper, a novel hysteresis compensator with learning-based pose estimation is proposed. Unlike previous studies, the proposed compensator can reduce hysteresis of the surgical instrument in various sheath configurations without using markers. In order to estimate an actual angle of the surgical instrument's joint, we employ the learning-based pose estimation using a siamese convolutional neural network (SCNN). The proposed compensator reduces hysteresis by partially controlling the position command, similar to the instinctive adjustments that physicians make with their visual feedback. To validate the proposed method, a testbed was constructed considering several requirements of flexible surgery robots. As a result, the results show the proposed method reduces hysteresis to less than 10°, for various configurations of sheath. In addition, we confirmed that the learning-based pose estimation is sufficient to apply to the proposed compensator for reducing hysteresis in real-time.

Laser tattoo removal is a standard non-invasive method for removing color pigments on the skin. Increasing number of tattooed people who want to remove their tattoo has driven the medical laser market to develop new technologies for painless, scar-free and complete tattoo removal. However, manual use of such laser systems creates post-operative complications since they do not guarantee (i) protection on nontattooed skin from laser exposure, nor (ii) precise control of the laser focus during the operations for best performance. This paper introduces deTattoo, a robotic system to improve tattoo removal operations. A robotic arm is equipped with a RGB-D camera and a visible laser, in eye-in-end configuration. A hybrid visual servoing control is proposed to guarantee the correct pose of the laser with respect to the tattooed tissue while compensating body motions. 2D features tracked with a mass-spring-damper deformable mesh model are combined with the 3D reconstruction retrieved from a RGB-D camera in order to build the control law. Several experiments were conducted to evaluate the performance of the system with a fixed or moving tattooed surface, at different inclinations. Results showed that the proposed framework is able to fulfil the laser-based tattoo removal requirements, providing high positioning accuracy (<; 1mm) orientation (<; 0.2°) and body motion compensation.

Contact-based operation in moving process is a challenging problem for aerial manipulators. It requires the whole system to maintain steady contact with external environment, to track some predefined trajectories on surfaces, and simultaneously to present some fixed contact force. Aiming at this problem, a hybrid force/motion control framework is proposed in this paper. In this framework, contact force control and position control are performed separately in two orthogonal subspaces: constrained space and free-flight space. To control the contact force, the closed-loop unmanned aerial vehicle is first theoretically shown to behave dynamically as a spring-mass-damper system. Further, an inverse-dynamics-based controller is proposed. To control the moving along the contact surface, trajectory planning and position controller are combined to achieve the steady behavior in a free-flight subspace. In the end, an aerial manipulator system with a roller-type end-effector was designed, and practical flight experiments was performed. The results indicate that the proposed framework is effective and validity.

We present a torque controller for humanoid robot's arm and a method to execute heavy-load tasks under that controller. Torque control of arms can reduce the joint load when an impulsive force is applied to the robot's hand. This feature is important for robots that will work among many humans or obstacles because accidental collisions may happen in such situations. We also developed a static force filter for force sensors at the end-effectors. This filter is utilized for the compensation for the static reaction force during heavy-load tasks. A life-sized humanoid robot JAXON digs soil with a shovel and carries soil with a wheelbarrow using our proposed controller.

This work presents an approach for humanoid Next Best View (NBV) planning that exploits full body motions to observe objects occluded by obstacles. The task is to explore a given region of interest in an initially unknown environment. The robot is equipped with a depth sensor, and it can perform both 2D and 3D mapping. As main contribution with respect to previous work, the proposed method does not rely on simple motions of the head and it was evaluated in real environments. The robot is guided by two behaviors: a target behavior that aims at observing the region of interest by exploiting body movements primitives, and an exploration behavior that aims at observing other unknown areas. Experiments show that the humanoid is able to peer around obstacles to reach a favourable point of view. Moreover, the proposed approach results in a more complete reconstruction of objects than a conventional algorithm that only changes the orientation of the head.

Robotic minimally invasive surgery has been a subject of intense research and development over the last three decades, due to the clinical advantages it holds for patients and doctors alike. Particularly for drug delivery mechanisms, higher precision and the ability to follow complex trajectories in three dimensions (3D), has led to interest in flexible, steerable needles such as the programmable bevel-tip needle (PBN). Steering in 3D, however, holds practical challenges for surgeons, as interfaces are traditionally designed for straight line paths. This work presents a pilot study undertaken to evaluate a novel human-machine visual interface for the steering of a robotic PBN, where both qualitative evaluation of the interface and quantitative evaluation of the performance of the subjects in following a 3D path are measured. A series of needle insertions are performed in phantom tissue (gelatin) by the experiment subjects. User could adequately use the system with little training and low workload, and reach the target point at the end of the path with millimeter range accuracy.

With recent development of robotic technology, it is increasingly common that robot coexist with human, in which humans and robots share a common workspace and work in close proximity. To maintain efficiency and ensure safety under these circumstances, robot should have the ability to predict the future human motion based on the observed on-going motion. In this paper, we present a methodology for on-line inference of human intention and prediction of human hand motion. The proposed framework is built using Probabilistic Dynamic Movement Primitive (PDMP). In the off-line stage, a set of PDMPs is constructed based on the recorded demonstrations and they will then be used for inferring human intention and predicting human hand motion in the on-line stage. A proof of concept evaluation is carried out in a tabletop manipulation task. Experimental result shows the proposed framework achieve high performance in human intention inference and in the trajectory similarity between the predicted and the actual hand movement under the normally defined environment. We also show the proposed framework can adapt and generalize to the newly defined environment.

Research focuses on robots related to area coverage applications such as cleaning, painting, demining, lawn moving, and inspection is gaining significant momentum in recent years. The majority of such research platforms faces significant performance issues while accessing narrow and constrained spaces due to their fixed morphology. To this end, we have developed a novel self-reconfigurable robot platform named as “hTetran” inspired from Polyabolo which can change its morphology to any of the six tetrabolo shapes with an objective of maximizing the area coverage. This paper presents the system design, reconfiguration theory, and locomotion modules of the developed robot, including the adaptation of Polyabolo tiling theory as a coverage path planning strategy for autonomous navigation. The paper concludes with a set of experiments in a mocked office room setup that validates the efficiency of the proposed robot in terms of area coverage. Our experimental trials indicate that the “hTetran” brings out a higher area coverage performance in all considered experimental cases.