After years of development, robots have become increasingly intelligent, resembling humans in many ways. The goal is to create new types of intelligent robots with human-like learning, motor, executive, perceptual, and cognitive abilities. Teaching-based learning is a crucial branch of research on robot learning capabilities and is key to the development and application of intelligent humanoid robots. Therefore, research on teaching-based learning has always been a hot topic in academia.
In recent years, to enable robots to accurately and quickly complete actions taught by human instructors, many teaching methods have emerged. For example, the KHR series robots produced by Kondo in Japan use direct input of the robot's joint angles for teaching. Devol in the United States proposed using servo technology to control the robot's joints, with a human hand teaching the robot the actions, and the robot recording and reproducing the teaching data. Cao Qixin et al. proposed using laser depth sensors to obtain the three-dimensional coordinates of the human body's joint angles, inputting them into a controller for teaching, and using a speed-based selective mean filtering method to process the control data. These teaching methods have made significant contributions to the research on human-robot teaching and have played a major role in promoting it. However, the accuracy of the robot's completed teaching actions, the speed of reproduction, and the control of sensor data jitter still need improvement.
In terms of system control strategies, many algorithms have been applied to robot teaching, such as Gaussian mixture models, reinforcement learning, and trial-and-error learning. While these methods have demonstrated good system stability and effectiveness in system control, they also have their own drawbacks, such as long training times and poor continuity of generated trajectories. Extreme Learning Machine (ELM) is a novel algorithm designed to address the shortcomings of the aforementioned algorithms, offering advantages such as strong generalization ability and fast training speed.
In summary, to ensure the robot teaching system possesses good adaptability and robustness—that is, the robot can return to the endpoint position according to the taught trajectory from different starting points, thereby giving the robot a certain degree of intelligence and self-learning ability—and to address the shortcomings of the aforementioned algorithms, this paper selects the NAO teaching and research robot as the experimental platform. Teaching is conducted by manipulating its arm joints. A robot teaching system based on the Extreme Learning Machine (ELM) algorithm is constructed, and experiments verify that this algorithm enables the system to have a certain degree of generalization ability.
Overview of Control Process
The ability for a robot to perform point-to-point teaching actions is a crucial component of humanoid robot motion editing research. Therefore, this paper uses the robot's movement from a starting point to a target point as the basic action. Sampling is performed at extremely short time intervals, collecting the values of the four joint angles on the NAO's right arm during each sampling. The data is then preprocessed, followed by regression training. The trained model is used to control the robot's controller, enabling the robot to reproduce the taught actions more accurately. (See Figure 1)
Introduction to the Experimental Platform
The small humanoid robot NAO, developed by the French company Aldebaran Robotics, is 58 centimeters tall and has 25 degrees of freedom. It possesses kinesthetic, visual, tactile, sensory, and cognitive abilities. It has a sophisticated digital processor and motor actuators, a rich sensor system, various communication devices, a unique operating system NAOqi, and a complete programming platform. It supports both Wi-Fi and Ethernet communication. Due to its excellent secondary development capabilities, it is widely used as a research tool in teaching, scientific research, and other fields.
Teaching learning algorithm
Because sensor measurements suffer from cumulative errors, noise, and instability, the measured data can be inaccurate. Therefore, a discrete-time Kalman filter (KF) is used to fuse the sensor output data from the NAO arm joints, optimizing the robot arm's posture accuracy and improving the accuracy of robot teaching and learning. The measurement values for the next time step are estimated based on the joint sensor values collected at the previous moment, and then fused with these values to correct the next moment's measurement values, thus making the data more accurate.
Based on the robot's actual motion state, the motion system is defined as a first-order ordinary differential equation:
experiment
HCS (Learning Human by Demonstration) is a statistical optimization algorithm proposed by Y. Ou and Y. Xu [12] to learn the parameters of a dynamic system by using support vector basis as a model in a nonlinear dynamic system, so as to ensure that the action can reach and stop at the target point according to the taught action. In order to verify that the ELM algorithm has the characteristics of convergence, fast training speed and high accuracy under the above constraints, the simulation experimental results of ELM and SEDS are compared firstly under the premise of using the same sample data. The sample data comes from a database of various human handwriting. Taking Cshape, Spoon, Line, Trapezoid and other data in the database as examples, the experimental comparison diagram is shown in Figure 3.
In a practical teaching experiment conducted on the experimental platform, the operator first manipulated the end effector of the NAO robot arm to perform the required obstacle avoidance and object retrieval actions. Simultaneously, the remote controller within the NAO recorded sensor and joint angle values every 100ms. When the robot needed to reproduce the taught action, the dynamic system retrieved the control information trained using the ELM algorithm from memory and transmitted the command signal to the drive mechanism, enabling the drive mechanism to accurately and quickly complete the required action. In the ELM, the number of hidden layer nodes was 50, and both the input and hidden layer weights were random vectors. The processed spatial position data of the robot arm end effector was used as the input to the Extreme Learning Machine, and the velocity of the end effector was used as its output.
The teaching process and teaching curve of the human mentor are shown in Figure 4.
Figure 5 shows the trajectory of the robot during obstacle avoidance teaching and reproduction from different starting points. The black dot set represents the starting point of the action, the blue dot set represents the ending point of the action, the red curve is the teaching curve composed of sample data, and the blue curve is the curve of the robot reproducing the taught action.
in conclusion
This paper combines the advantages of the ELM learning algorithm, such as fast training speed and few parameters requiring adjustment, to teach robots, overcoming the shortcomings of previous humanoid robot teaching methods, such as slow learning speed and low accuracy in reproducing taught actions. Practical results show that this method has good anti-interference ability and good generalization ability, while avoiding the curse of dimensionality. The ELM-based dynamic system is described as a nonlinear dynamic system using ordinary differential equations, and the constraints for the system to achieve local stability at the target point are given. Lyapunov's stability theorem provides theoretical support for its stability, laying the foundation for future humanoid robots to learn more complex taught actions.
For more information, please follow the Sensor Channel.
