Squeezing More Past Knowledge for Online Class-Incremental Continual Learning

Continual learning(CL)studies the problem of learning to accumulate knowledge over time from a stream of data.A crucial challenge is that neural networks suffer from performance degradation on previously seen data,known as catastrophic forgetting,due to allowing parameter sharing.In this work,we consider a more practical online class-incremental CL setting,where the model learns new samples in an online manner and may continuously experience new classes.Moreover,prior knowledge is unavailable during training and evaluation.Existing works usually explore sample usages from a single dimension,which ignores a lot of valuable supervisory information.To better tackle the setting,we propose a novel replay-based CL method,which leverages multi-level representations produced by the intermediate process of training samples for replay and strengthens supervision to consolidate previous knowledge.Specifically,besides the previous raw samples,we store the corresponding logits and features in the memory.Furthermore,to imitate the prediction of the past model,we construct extra constraints by leveraging multi-level information stored in the memory.With the same number of samples for replay,our method can use more past knowledge to prevent interference.We conduct extensive evaluations on several popular CL datasets,and experiments show that our method consistently outperforms state-of-the-art methods with various sizes of episodic memory.We further provide a detailed analysis of these results and demonstrate that our method is more viable in practical scenarios.

Reinforcement Learning Navigation for Robots Based on Hippocampus Episode Cognition

Artificial intelligence is currently achieving impressive success in all fields.However,autonomous navigation remains a major challenge for AI.Reinforcement learning is used for target navigation to simulate the interaction between the brain and the environment at the behavioral level,but the Artificial Neural Network trained by reinforcement learning cannot match the autonomous mobility of humans and animals.The hippocampus–striatum circuits are considered as key circuits for target navigation planning and decision-making.This paper aims to construct a bionic navigation model of reinforcement learning corresponding to the nervous system to improve the autonomous navigation performance of the robot.The ventral striatum is considered to be the behavioral evaluation region,and the hippocampal–striatum circuit constitutes the position–reward association.In this paper,a set of episode cognition and reinforcement learning system simulating the mechanism of hippocampus and ventral striatum is constructed,which is used to provide target guidance for the robot to perform autonomous tasks.Compared with traditional methods,this system reflects the high efficiency of learning and better Environmental Adaptability.Our research is an exploration of the intersection and fusion of artificial intelligence and neuroscience,which is conducive to the development of artificial intelligence and the understanding of the nervous system.

Robot Navigation Strategy in Complex Environment Based on Episode Cognition

The hippocampal formation of the brain contains a series of nerve cells related to environmental cognition and navigation.These cells can integrate their moment information and external perceptual information and acquire episodic cognitive memory.Through episodic cognition and memory,organisms can achieve autonomous navigation in complex environments.This paper mainly studies the strategy of robot episode navigation in complex environments.After exploring the environment,the robot obtains subjective environmental cognition and forms a cognition map.The grid cells information contained in the cognitive map can obtain the direction and distance of the target through vector calculation,which can get a shortcut through the inexperienced area.The synaptic connection of place cells in the cognitive map can be used as the topological relationship between episode nodes.When the target-oriented vector navigation encounters obstacles,the obstacles can be realized by setting closer sub-targets.Based on the known obstacle information obtained from boundary cells in the cognitive map,topological paths can be divided into multi-segment vector navigation to avoid encountering obstacles.This paper combines vector and topological navigation to achieve goal-oriented and robust navigation capability in a complex environment.

CPG Control for Biped Hopping Robot in Unpredictable Environment

A CPG control mechanism is proposed for hopping motion control of biped robot in unpredictable environment. Based on analysis of robot motion and biological observation of animal＇s control mechanism, the motion control task is divided into two simple parts： motion sequence control and output force control. Inspired by a two-level CPG model, a two-level CPG control mechanism is constructed to coordinate the drivers of robot joint, while various feedback information are introduced into the control mechanism. Interneurons within the control mechanism are modeled to generate motion rhythm and pattern promptly for motion sequence control; motoneurons are modeled to control output forces of joint drivers in real time according to feedbacks. The control system can perceive changes caused by unknown perturbations and environment changes according to feedback information, and adapt to unpredictable environment by adjusting outputs of neurons. The control mechanism is applied to a biped hopping robot in unpredictable environment on simulation platform, and stable adaptive motions are obtained.

A Parallel Actuated Pantograph Leg for High-speed Locomotion

High-speed running is one of the most important topics in the field of legged robots which requires strict constraints on structural design and control. To solve the problems of high acceleration, high energy consumption, high pace frequency and ground impact during high-speed movement, this paper presents a parallel actuated pantograph leg with an approximately decoupled configuration. The articulated leg features in light weight, high load capacity, high mechanical efficiency and structural stability. The similarity features of force and position between the control point and the foot are analyzed. The key design parameters, K1 and K2, which concern the dynamic performances, are carefully optimized by comprehensive evaluation of the leg inertia and mass within the maximum foot trajectory, A control strategy that incorporates virtual Spring Loaded Inverted Pendulum （SLIP） model and active force is also proposed to test the design. The strategy can implement highly flexible impedance without mechanical springs, which substantially simplifies the design and satisfies the variable stiffness requirements during high-speed running. The rationality of the structure and the effectiveness of the control law are validated by simulation and experiments.

Force-controlled Compensation Scheme for P-Q Valve-controlled Asymmetric Cylinder used on Hydraulic Quadruped Robots

Under the requirement of the force controller of hydraulic quadruped robots,the goal of this work is to accurately track the force commands at the level of the hydraulic drive unit.The main contribution focuses on the development of a force-controlled compensation scheme,which is specifically aimed at the key issues affecting the hydraulic quadrupedal locomotion.With this idea,based on a P-Q valve-controlled asymmetric cylinder,we first establish a mathematical model for the hydraulic drive unit force control system.With the desired force commands,a force feed-forward algorithm is presented to improve the dynamic performance of the system.Meanwhile,we propose a disturbance compensation algorithm to reduce the influence induced by external disturbances due to foot-ground impacts.Afterwards,combining with a variable gain PI controller,a series of experiments are implemented on a force control performance test platform to verify the proposed scheme.The results demonstrate that the force-controlled compensation scheme has the ability to notably improve the force tracking accuracy,reduce the response time and redundant force.