Unmanned Systems

This paper concentrates on the disturbances estimation and attenuation problem of a general class of the fractional-order systems. For this purpose, this paper proposes a fractional sliding-mode disturbance observer (FSMDO)-based control scheme, which is capable of mitigating the unknown disturbances asymptotically. Meanwhile, by incorporating a novel fractional sliding-mode controller into the proposed control scheme, the asymptotical convergence of trajectory tracking errors is guaranteed. The developed control scheme owns three attractive highlights: (1) it is suitable for the general fractional-order systems, including nonlinear systems and incommensurate systems; (2) the proposed FSMDO can estimate the unknown exogenous disturbances and system uncertainties timely and precisely; (3) nominal performance can be retained by compensating the observed disturbances in a feedforward manner, and thus, the robustness of the system can be strengthened. Illustrative examples are provided to show the availability and superiority of the presented FSMDO control method in terms of robust control. As compared with the conventional methods, the dynamic control performance of the closed-loop system can be improved.

Though the traditional algorithms could be embedded into neural architectures with the proposed principle of [H. Xiao, Hungarian layer: Logics empowered neural architecture, arXiv: 1712.02555], the variables that only occur in the condition of branch could not be updated as a special case. To tackle this issue, we multiply the conditioned branches with Dirac symbol (i.e., $1_{x>0}$), then approximate Dirac symbol with the continuous functions (e.g., $1-e^{-\alpha |x|}$). In this way, the gradients of condition-specific variables could be worked out in the back-propagation process, approximately, making a fully functional neural graph. Within our novel principle, we propose the neural decision tree (NDT), which takes simplified neural networks as decision function in each branch and employs complex neural networks to generate the output in each leaf. Extensive experiments verify our theoretical analysis and demonstrate the effectiveness of our model.

Dexterous manipulation of dual-arm robots in unstructured environments is very useful. Programming a dual-arm industrial robot to efficiently complete dexterous tasks, however, is especially challenging due to the complexity of its inverse kinematics, motion planning, dual-arm coordination with self-collision avoidance, and so on. This paper presents a systematic solution to accurately manipulate a dual-arm industrial robot on-site via light-weight wearable devices. In the developed system, the human operator directly drives the robot through the human arms motions tracked by the combination of inertial measurement units and handheld joystick controllers. A proper motion retargeting method with self-collision avoidance is used to enable the user to manipulate the robot directly through intuitive arm motions within a comfortable range and ensure the task manipulation with safety in unstructured environments. The developed system has been tested with various tasks, such as the manipulation of objects of different shapes, dexterous turn-over, and dual-arm coordination. Compared with the existing telerobotic systems, the developed system with simultaneous 14 degree-of-freedom teleoperation directly driven by light-weight wearable devices is able to handle more dexterous and accurate manipulation tasks with the capability of fast deployment and self-collision awareness. Such a solution could pave the way for online dual-arm robot programming on efficient manipulation skills transfer in the future.

Adaptive structures are able to react to environmental impacts and have become a promising approach in civil engineering to improve the load-bearing behavior of buildings. Since reliability and safety of building structures are major concerns, the detection and isolation of faults are essential. In this work, the data-based distributed fault diagnosis of sensor and actuator faults in an adaptive high-rise truss structure is investigated and compared to a centralized approach. The decomposition of the different subsystems is given by the hardware layout of the different sensor systems and actuators. The mechanical structure is modeled and extended by dynamic sensor and actuator models containing different faults. Based on the simulation model, different fault scenarios are generated and used for training a convolutional neural network with dropout regularization. It is shown that the distributed approach needs less training data and yields better classification results than the centralized approach due to a significant reduction of the complexity and dimensionality.

The combination of data from multiple sensors, also known as sensor fusion or data fusion, is a key aspect in the design of autonomous robots. In particular, algorithms able to accommodate sensor fusion techniques enable increased accuracy, and are more resilient against the malfunction of individual sensors. The development of algorithms for autonomous navigation, mapping and localization have seen big advancements over the past two decades. Nonetheless, challenges remain in developing robust solutions for accurate localization in dense urban environments, where the so-called last-mile delivery occurs. In these scenarios, local motion estimation is combined with the matching of real-time data with a detailed pre-built map. In this paper, we utilize data gathered with an autonomous delivery robot to compare different sensor fusion techniques and evaluate which are the algorithms providing the highest accuracy depending on the environment. The techniques we analyze and propose in this paper utilize 3D lidar data, inertial data, GNSS data and wheel encoder readings. We show how lidar scan matching combined with other sensor data can be used to increase the accuracy of the robot localization and, in consequence, its navigation. Moreover, we propose a strategy to reduce the impact on navigation performance when a change in the environment renders map data invalid or part of the available map is corrupted.

This paper proposes a compliant motion control for human-cooperative robots to absorb collision force when persons accidentally touch the robots even while the robot is manipulating an object. In the proposed method, an external force estimator, which can distinguish the net external force from the object manipulation force, is realized using an inverse dynamics model acquired by a recurrent neural network (RNN). By implementing a mechanical impedance control to the estimated external force, the robot can quickly and precisely carry the object keeping the mechanical impedance control functioned and can generate a compliant motion to the net external force only when the person touches it during manipulation. Since the proposed method estimates the external force from the generalized force based on the learned inverse dynamics, it is not necessary to install any sensors on the manipulated object to measure the external force. This allows the robot to detect the collision even when the person touches anywhere on the manipulated object. The RNN inverse dynamics model is evaluated by the leave-one-out cross-validation and it was found that it works well for unknown trajectories excluded from the learning process. Although the details were omitted due to the limitation of the page length, similar to the simulations, the RNN inverse dynamics model was evaluated using unknown trajectories in the six degree-of-freedom experiments, and it has been verified that it functions properly even for the unknown trajectories. Finally, the validity of the proposed method has been confirmed by experiments in which a person touches a robot while it is manipulating an object with six degrees of freedom.

This work develops and implements a multi-agent time-based path-planning method using A*. The purpose of this work is to create methods in which multi-agent systems can coordinate actions and complete them at the same time. We utilized A* with constraints defined by a dynamic model of each agent. The model for each agent is updated during each time step and the resulting control is determined. This results in a translational path that each of the agents is physically capable of completing in synchrony. The resulting path is given to the agents as a sequence of waypoints. Periodic updates of the path are calculated, utilizing real-world position and velocity information, as the agents complete the task to account for external disturbances. Our methodology is tested in a dynamic simulation environment as well as on real-world lighter-than-air robotic agents.