Online Continual Learning (OCL), as a core step towards achieving human-level intelligence, aims to incrementally learn and accumulate novel concepts from streaming data that can be seen only once, while alleviating catastrophic forgetting on previously acquired knowledge. Under this mode, the model needs to learn new classes or tasks in an online manner, and the data distribution may change over time. Moreover, task boundaries and identities are not available during training and evaluation. To balance the stability and plasticity of networks, in this work, we propose a replay-based framework for OCL, named Contrastive Correlation Preserving Replay (CCPR), which focuses on not only instances but also correlations between multiple instances. Specifically, besides the previous raw samples, the corresponding representations are stored in the memory and used to construct correlations for the past and the current model. To better capture correlation and higher-order dependencies, we maximize the low bound of mutual information between the past correlation and the current correlation by leveraging contrastive objectives. Furthermore, to improve the performance, we propose a new memory update strategy, which simultaneously encourages the balance and diversity of samples within the memory. With limited memory slots, it allows less redundant and more representative samples for later replay. We conduct extensive evaluations on several popular CL datasets, and experiments show that our method consistently outperforms the state-of-the-art methods and can effectively consolidate knowledge to alleviate forgetting.

A robot in the future may initially has a good learning capability but an empty library of movements. It gradually enriches its library of movements through human demonstrations. Dynamic Movement Primitives (DMPs) has been proved to be an effective way to represent trajectories. Trajectories are classified into discrete and rhythmic ones, and parameters are set for each demonstrated trajectory. However, what kind of trajectory will be provided by robot users is sometimes unknown to robot developers, so trajectory pattern and the parameters can not be determined in advance. It's also impossible for non-technical robot users to set these parameters and determine the pattern of movements they are going to demonstrate. To make it easier for non-expert robot users to programme their robots by demonstration, this work presents an efficient way to deal with these two problems. The effectiveness of the proposed methodology is proved by teaching a robot to clean the whiteboard in different ways and stack a set of cubic boxes in specific order.

Designing an e-commerce recommender system that serves hundreds of millions of active users is a daunting challenge. To our best knowledge, the complex brain activity mechanism behind human shopping activities is never considered in existing recommender systems. From a human vision perspective, we found two key factors that affect users’ behaviors: items’ attractiveness and their matching degrees with users’ interests. This paper proposes Telepath, a vision-based bionic recommender system model, which simulates human brain activities in decision making of shopping, thus understanding users from such perspective. The core of Telepath is a complex deep neural network with multiple subnetworks. In practice, the Telepath model has been launched to JD’s recommender system and advertising system and outperformed the former state-of-the-art method. For one of the major item recommendation blocks on the JD app, click-through rate (CTR), gross merchandise value (GMV) and orders have been increased 1.59%, 8.16% and 8.71% respectively by Telepath. For several major ad publishers of JD demand-side platform, CTR, GMV and return on investment have been increased 6.58%, 61.72% and 65.57% respectively by the first launch of Telepath, and further increased 2.95%, 41.75% and 41.37% respectively by the second launch.

In this paper, a joint control algorism was proposed to implement on our musculoskeletal robot. The joints are all actuated by pneumatic muscles and have antagonistic structure. In order to gain a better performance, we first modeled the pneumatic muscle, and then considering the dynamics two control method was discussed and compared. As a tradeoff of rapidity and accuracy, the combined joint control algorism was implemented on the robot system. The algorism could tune the stiffness of the joint automatically to fit the compliance need for joint control. The experiments on our robot showed that the joint control algorism could operate the robot with a fast and accurate response. There is nearly not lagging behind, and with a small overshot during the step testing, and a complex leg's trajectory was also achieved through our joint controller.

This paper proposes a novel CPG-based control architecture for hexapod walking robot. We investigate the CPG systems from the perspective of network synchronization. In this way the motion control of hexapod robot can be refined into the gait generation level and joints coordination level. On the first level, we develop a gait generator consists of CPG network in ring based on modified Van der Pol (VDP) oscillator to realize various stable gaits as well as gait transition for hexapod walking. The limit cycle behavior of VDP model is analytically studied by virtue of perturbation technique. On the second level, we address the problem of multi-DoF coordination of single leg via phase order modulation and amplitude adjustment of the neural oscillators. Consequently we propose a single-leg controller consists of a three-coupled CPG network and a linear coefficient converter to generator smooth and feasible trajectories in task space. The effectiveness of the proposed control architecture is demonstrated through simulation and real physical robot experiment.

The Spring-Loaded Inverted Pendulum (SLIP), or monopedal runner, is widely used to depict running and hopping in mammalian and human locomotion, which is also serving as a template for running robot design. This classic model describes quite a simple mechanical system. Nevertheless issue of seeking the accurate analytic solution revealing the characteristics of the motion during stance remains unsettled due to the nonintegrable terms contained in the system equations. Moreover, several existing analytic approximations by simply ignoring or linearizing the gravitational force can not reveal the entire dynamical behavior of nonlinear system as well as can be breakdown rapidly when applied to a non-symmetric motion case. In this paper, a novel method with perturbation technique is proposed to obtain analytic approximate solutions to the SLIP dynamics in stance phase with considering the effect of gravity. The perturbation solution achieves higher accuracy in predicting the apex trajectory and stance locomotion by comparing with typical existing analytical approximations. Particularly, our solution is validated for non-symmetric case in a large angle range. Additionally, the prediction for stance trajectory is also verified through numerical evaluation.

The high speed running of legged robot raise higher requirements on the design and development of the robot and the control method. The mammals' excellent performance, especially the cheetah, provides us much inspiration on control, mechanical design and etc. In this paper, a model of the cheetah robot is developed for researching the locomotory behaviors of fast running. Under the biological knowledge on neural mechanism, a new bio-inspired control method was proposed via controlling the muscles flexion and extension movements. Then, a bio-musculoskeletal hindlimb prototype was designed in this paper. The muscle-like actuator and the cheetah-like structure combining the bio-control approach provide more natural fashion for the new robot. Finally, the simulation of the bounding gait establishes good biological properties, cheetah-like leg's trajectory. Experiment with the real hindlimb prototype validates the design and the development of the structure and the controller. In the near future, a final prototype will be manufactured with onboard power and air source.

In this paper, we introduce a quadruped robot designed for bounding gait with only one actuator per compliant leg. Under the analysis of the dynamics model of the robot, a new simple linear running controller using the energy control idea, which requires minimal task level feedback and only controls both the leg torque and ending impact angle, is proposed. It successfully executes fast running from rest till a constant speed and hi-speed turning, both in the prototype simulation and robot experiment. These results contribute to that complex dynamically dexterous tasks may be controlled via simple energy control method and delayed task feedback, which is closer to the animal's actual locomotion conditions. In the future, we plan to modify this method for reducing the energy expending and make the robot running fast.

This paper deals with the dynamics and motion control of a spherical robot designed for reconnaissance and unstructured hostile environment exploration. The robot in this paper has three DOFs and two inputs, of which the nature is a nonlinear and underactuated system with nonholonomic dynamic constraints. The improved construction of two pendulums offers novel motion principle of spherical robot, which is moving simultaneously actuated by both eccentric moment and inertial moment generated by the two pendulums. Meanwhile the mobility is enhanced when the robot behaves dynamically. The emphasis is placed on the linear motion and turning in place motion control. The dynamic model of linear motion is formulated on the basis of Lagrange equation, and a smooth trajectory planning method is proposed for linear motion. A feedback controller is constructed to ensure the accurate trajectory planning. Turning in place motion is an indispensable element of omni directional locomotion which can enhance the mobility of spherical robots. The dynamic model of turning in place motion is derived on the theory of moment of momentum, and a stick-slip principle is analyzed. The two motion control methods are validated by both simulations and prototype experiments.

Robots become more and more widely used in medical application especially in minimally invasive surgery (MIS) these years, and endoscope robot is a typical example. There is sill not a kind of conventional endoscope which can reach the whole gastrointestinal (GI) tract. And conventional endoscopes generate pain and discomfort to patients due to the stiffness of their body. Such disadvantages do not exist in endoscope robots. However, the soft and elastic characters of GI tract make the locomotion efficiency of endoscope robot very low. To resolve this problem, it is necessary to find a special locomotion principle which fit the GI tract, and many researchers have paid a lot of effort. In this paper, a novel locomotion principle is proposed, which is the first time used in the endoscope robot. The locomotion principle is obtained by analyzing, simulating and simplifying the movement of mucus-cilia system. To evaluate its performance, a prototype of the endoscope robot adopting this locomotion principle is developed, which is driven by shape memory alloy (SMA) actuator. An in-vitro experiment shows that the robot can move successfully with efficiency about 27%. The result of experiment validates that the proposed locomotion principle fit the GI tract. And some suggestions to enhance the locomotion efficiency are proposed

Pixel-based auto-focusing is a long-standing topic in the literatures. It involves three main parameters: Region of Interest (ROI), image sharpness function and global maximum searching algorithm. As the mathematical description of image sharpness, sharpness function is the core issue for realizing robust auto-focusing. In this paper, the existing sharpness functions are summarized and grouped firstly, then a new space domain SUSAN (Smallest Univalue Segment Assimilating Nucleus) based sharpness function is proposed. The key problem in proposed algorithm is selecting a suitable similarity function to measure the similarities of the sampling pixel and its neighbors. In experiments, 8 similarity functions are analyzed and evaluated. Based on the evaluation results, a rough/fine two grades global maximum searching strategy is designed to realize fast and robust auto-focusing. At last, experiments verify the validity of the proposed auto-focusing algorithm.

Micro operation assisted by miniature mobile robot system has become a research hot in robotics and micro operation domain. In this paper, a newly developed micro operation oriented five D.O.F. (degrees of freedom) miniature mobile robot system is presented. By comparing it with other miniature robot systems, the main improvements of this design are macro/micro dual driven, embedded hardware, wireless power supply and wireless communication Hereinto, development of a mobile robot with dual motions is the first challenge and the main original contribution. Two miniaturized brushless DC motors drive the macro motion, which realizes faster motion of the robot. Micro motion, which realizes precision positioning of the robot, is driven by four piezoelectric-actuators and four electromagnetic foots based on inchworm principle. To guide the robot accomplishing micro operation automatically, an external visual sensor system with CCD cameras and optical zoom microscopes is designed. And the details of robot control architecture are explained. Experiments show that the proposed robot system succeeds in providing micro operation task such as micro part gripping, convey and assembly.