We propose an efficient spectral clustering method for large-scale data. The main idea in our method consists of employing random Fourier features to explicitly represent data in kernel space. The complexity of spectral clustering thus is shown lower than existing Nyström approximations on large-scale data. With m training points from a total of n data points, Nyström method requires O(nmd+m³+nm²) operations, where d is the input dimension. In contrast, our proposed method requires O(nDd+D³+n'D²), where n' is the number of data points needed until convergence and D is the kernel mapped dimension. In large-scale datasets where n' << n hold true, our explicitly mapping method can significantly speed up eigenvector approximation and benefit prediction speed in spectral clustering. For instance, on MNIST (60,000 data points), the proposed method is similar in clustering accuracy to Nyström methods while its speed is twice as fast as Nyström.

Accurate position-force control is a core and challenging problem in robotics, especially for manipulators with redundant DOFs. For example, trajectory tracking based control usually fails for grinding robots due to intolerable impact forces imposed onto the end-effectors. The main difficulties lie in the coupling of motion and contact force, redundancy resolution and physical constraints, etc. In this paper, we propose a novel motionforce control strategy in the framework of projection recurrent neural networks. Tracking error and contact force are described in orthogonal spaces respectively, and by selecting minimizing joint torque as secondary task, the control problem is formulated as a quadratic-programming (QP) problem under multiple constraints. In order to obtain real-time optimization of joint toque which is non-convex relative to joint angles, the original QP is reconstructed in velocity level, where the original objective function is replaced by its time derivative. Then a dynamic neural network which is convergence provable is established to solve the modified QP problem online. This work generalizes projection recurrent neural network based position control of manipulators to that of position-force control, which opens a new avenue to shift position-force control of manipulators from pure control perspective to cross design with both convergence and optimality consideration. Numerical and experimental results show that the proposed scheme achieves accurate position-force control, and is capable of handling inequality constraints such as joint angular, velocity and torque limitations, simultaneously, consumption of joint torque can be decreased effectively.

To promote operational intelligence, improve surface quality, and reduce manpower dependence, a novel high-bandwidth end-effector with active force control for robotic polishing was proposed. Using this end-effector as a mini robot, a macro-mini robot for polishing processing was constructed, in which the macro robot provides posture control during polishing operations, whereas the mini-robot provides constant force control. By minimizing the inertia along the spindle in this configuration, the end-effector obtains a force control bandwidth of 200 Hz. Through a series of comparative experiments with different contact forces and feed rates, the proposed design was proven to have a smaller overshoot, a faster response, and a shorter settling time than the conventional method based on macro robot (KUKA iiwa) controlled force. The roughness of the workpiece reached 0.4 μm after polishing with the macro-mini robot, indicating the efficiency of this end-effector in high-precision material removal and surface polishing operations.

Agriculture, forestry, and building industry would be prospective fields of robotic applications. High-rise tasks in these fields require robots with climbing skills. Motivated by these potential applications and inspired by animal climbing motion, we have developed a biped climbing robot—Climbot. Built with a modular approach, the robot consists of five joint modules connected in series and two special grippers mounted at the ends, with the scalability of changing degrees-of-freedom (DoFs). With this configuration, Climbot not only has superior mobility on multiple climbing media, such as poles and trusses, but also has the function of grasping and manipulating objects. It is a kind of “mobile” manipulator and represents an advancement in development of climbing robots. In this paper, we first present the development of this climbing robot with modular and bio-inspired methods, and then propose and compare three climbing gaits based on the unique configuration and features of the robot. A series of challenging and comprehensive experiments with the robot climbing in a truss and performing an outdoor manipulation task are carried out, to illustrate the feasibility, the features, the climbing, and manipulating functions of the robot, and to verify the effectiveness of the proposed gaits.

Deep Reinforcement Learning (RL) is a promising approach for adaptive robot control, but its current application to robotics is currently hindered by high sample requirements. To alleviate this issue, we propose to exploit the symmetries present in robotic tasks. Intuitively, symmetries from observed trajectories define transformations that leave the space of feasible RL trajectories invariant and can be used to generate new feasible trajectories, which could be used for training. Based on this data augmentation idea, we formulate a general framework, called Invariant Transform Experience Replay that we present with two techniques: (i) Kaleidoscope Experience Replay exploits reflectional symmetries and (ii) Goal-augmented Experience Replay which takes advantage of lax goal definitions. In the Fetch tasks from OpenAI Gym, our experimental results show significant increases in learning rates and success rates. Particularly, we attain a 13, 3, and 5 times speedup in the pushing, sliding, and pick-and-place tasks respectively in the multi-goal setting. Performance gains are also observed in similar tasks with obstacles and we successfully deployed a trained policy on a real Baxter robot. Our work demonstrates that invariant transformations on RL trajectories are a promising methodology to speedup learning in deep RL.

Industrial robots are increasingly being considered for application in performing polishing operations. Most research has been performed on workpieces with a priori (CAD) model, and used impedance control or hybrid force-position control to achieve constant force operation. For workpieces without priori models, it is still difficult to maintain a constant normal contact force during the polishing process. To address this challenge, this paper proposes a cooperative force-position control method consisting of three parts: constructing a model between contact forces and robotic poses during polishing operations and obtaining a real-time pose deviation of the current tool-workpiece contact state and the desired (normal) contact state; based on the deviation, a method of real-time adjustment of 2D predefined polishing paths to 3D actual polishing paths is proposed; finally, constant normal contact force control is achieved by a direct force control algorithm based on the results obtained from the above method. The proposed method is based on a macro-mini robotic system with an active force control polishing device, and its effectiveness was tested by various unknown-model workpiece polishing scenarios. The results demonstrate that the proposed method can achieve stable normal force tracking and high-quality surfaces during unknown-model workpiece polishing.