In recent years, robots have become more and more common helpers in people’s families. Four scientists from Carnegie Mellon University (CMU) found that there is a big gap between the application of home robots in real life and experiments. In the paper, they emphasized that robots generally live in laboratories, and life scenes are almost zero in real scenes.
Why doesn't robot motion data get similar benefits as we have seen in other important fields such as computer vision and natural language processing?
Many methods claim that the data collected in the laboratory is real data. However, in the face of diverse scenarios in reality, many robots seem to be powerless. Therefore, the data collection information of the robot during the experiment needs to be transferred from the laboratory setting to the home of people in the real world.
In response, scientists decided to take these robots to open houses. By physically executing instructions on the robot in multiple invisible homes, comparing the display of the model trained on the home data set with the baseline model trained on the data collected in the laboratory, to find a solution to the situation of the robot outside the laboratory data. Deal with and solve the problem of inaccurate response of some cheap robots to doing things in daily life and facing instructions.
Affordable robotSince the real-time cost of collecting simulation data is much lower than that of real-time robots, most of the current data-driven methods of robots mainly focus on the use of simulators, which do not use hand-designed models, but focus on large-scale data sets. Collect, but there is a wide "realistic gap" between the simulator and the real world. Therefore, scientists decided to promote the collection of real-world physical interaction data in multiple robotic laboratories, with the main purpose of reducing hardware costs.
For this purpose, the scientists modified the robot. The robot arm first built a low-cost mobile manipulator handle that can be assembled for less than 3K dollars, and then gradually added a dual-axis wrist, two-finger electric clamp, and a mobile base. In terms of sensors, it is equipped with an Intel R200 RGB camera and a gimbal that turns the camera around. As for the robot's brain is equipped with i5-8250U CPU and 8G RAM, it can run for about 3 hours on a single charge. The battery in the base is used to power the base and arm. With only one charge, the system can run for 1.5 hours. In this way, each "only" three thousand U.S. dollars, which is more economical than other (20,000 U.S. dollars) robots.
Because of the cheap electric motors, an inevitable consequence of lowering costs is highlighted-that is, the inability to accurately control. There are many errors in the various data collections of existing robots. This error is called the noise of the robot's work, and the noise is simulated as a latent variable and two networks are used: predict the possible noise and predict the action to be performed.
Robot plane grasping principleThe plane grasping training principle follows the imagenet pre-trained convolutional neural network as initialization, which is divided into 3 structures:
1. Grab Prediction Network (GPN), infer the grasp angle based on the image block of the object, and then decide which posture to grasp. With economical robots, the collected data will have a lot of noise
2. Noise modeling network (NMN), which estimates the two sets of potential noise scenes and robot information of a given image to separate the noise.
3. Marginalization Layer, which calculates the final grip angle and combines the two data streams in order to make better decisions.
Robot potential noise modelTo ensure diversified data testing, the scientists cited six families for plane capture. Each home has several environments, and multiple robots are used to collect data in parallel. Since the data is collected in a household with unstructured visual input, an object detector is used. This leads to the prediction of bounding boxes of objects in clutter and different backgrounds, so only the 2D position is recorded and the object class information is discarded. Once the position of the object is obtained in the image space, it is first grabbed and sampled, and then the 3D grabbed position is calculated from the noisy PointCloud.
Since the underconstrained robot has only 5 DOF, the motion planning pipeline is carefully designed. When collecting training data, various objects are scattered, so that the mobile base can move and grab the objects randomly. The base is restricted to a 2 meter wide area to prevent the robot from colliding with obstacles outside the operating area, and then quantitatively evaluate the collected data set.
For quantitative evaluation, three different test settings can be used:1. Binary classification (retained data): Collect and maintain test sets by performing random grabs on objects.
In the case of a given position and grasping angle, the performance of binary classification is measured. The model must predict whether the mastery is successful. This method allows to evaluate a large number of models without running them on a real robot.
2. Real low-cost arm (Real-LCA): By evaluating the physical grasping performance of the learning model on the low-cost arm.
3. Real Sawyer (Real-Sawyer): Measure the physical grasping performance of the learning model on the industrial robot arm (Sawyer). Since Sawyer is a more accurate and better calibration, the noise in the data will not be resolved during the evaluation of the Robust-Grasp model.
Robots cannot fully adapt to the outside world from model training to data set integration evaluation. Therefore, collecting real-world data to train the actual skills of the robot in real time is a generalized and difficult process. I am looking forward to the world of living with the robot. , Are you looking forward to it?
XB2-B Series Pushbutton Switch
XB2-B series Pushbutton Switch are using controlling circuit of AC voltage up to 660V/AC 50Hz.and DC voltage below 400V.It is used for controlling signal and interlocking purposes.The main body of the push buttons adopts Circuit head,with chromium plated metal bezel,Complete units with screw and captive cable clamp connections.The contact adopts special silver alloy in order to enjoy the benefit of high anti-electrical erosion.The signal lamp series are suitable for electrical applicances circuit of AC voltage up to 380V/50Hz,and DC voltage below 380V.It is ideal for using as indicating signals,warning signals,emergency signals.
Push Button On Off Switch,Waterproof Push Button Switch,Micro Push Button On Off Switch,Green Push Button Switch
Ningbo Bond Industrial Electric Co., Ltd. , https://www.bondelectro.com