For the first time, MIT researchers have been able to understand the configuration of the robot arm in 3D space by using only motion and position data from its "sensed" skin, which may be another change of the robot arm in the future.
According to the Internet technology media, this soft material is made of highly flexible materials (similar to those found in organisms)
Robot
It is also influenced by the idea that robots are more secure, more flexible, more flexible and adaptive than traditional rigid robots, and a new generation of robot substitutes with biological characteristics.
But the difficulty also exists! For example, autonomous control of these deformable robots is an arduous task, because these new software robots can move in almost infinite directions at any given time, which makes it difficult for researchers to train planning and control models for driving automatic devices through programming or teaching.
Previously, the traditional method of autonomous control (control system) was to use a large-scale vision system with multiple motion capture cameras, which can provide feedback on 3D motion and position for robots. However, at present, for the actual
application
For the new software robot in, this is impractical.
In a paper published in IEEE robotics and automation letters, researchers describe a software sensor system that covers the robot's body to provide "ontological feeling", that is, to sense the movement and position of its body. The feedback will enter a novel deep learning model, which can filter out noise and capture clear signals to calculate the 3D position of the robot. The researchers tested their system on a soft robot arm that looks like an oak trunk, allowing the robot arm to swing and stretch automatically and predict its spatial position.
picture
Source: Ryan L. Truby, MIT CSAIL
The researchers' soft sensors cut conductive silicon films into origami shapes, giving them "piezoresistive" properties, which means they change their resistance as they strain. When the sensor deforms in response to the tension and compression of the manipulator, its resistance is converted into an output voltage, which is then used as a signal related to the motion.
Ryan Truby, a postdoctoral fellow at the Massachusetts Institute of technology's computer science and artificial laboratory (CSAIL), said the sensor could be made of ready-made human materials, which means that any laboratory in the future can develop its own system.
"We are sensing the soft robot's characteristics to get feedback from sensors (rather than the visual system) for control, rather than through the visual system as before," he said "For example, we want to use these soft robot trunks to automatically orient and control ourselves, pick up things and interact with the world. This is the first step towards this more complex automation. "
One of the goals of the future is to help create artificial limbs that can handle and manipulate objects in the environment more nimbly. "Think of the human body: you can close your eyes and learn about the world based on skin feedback," said Daniel Rus, director of CSAIL and Professor Andrew and Erna Viterbi of the Department of electrical engineering and computer science, who co authored a book "We want to design the same functions for soft robots."
Shaping soft sensor
Image source: MIT
A fully integrated human body sensor is the long-term goal of soft robot technology. Traditional rigid sensors will damage the natural flexibility of soft robot, complicate its design and manufacture, and may lead to various mechanical failures. Therefore, sensor based on soft material is a more suitable alternative, but its design needs special material and program operation method, which makes it difficult for many robot laboratories to manufacture and integrate them in soft robot.
One day, while working in his CSAIL lab, looking for inspiration for sensor materials, Truby made interesting connections with these new materials. "I have found that these conductive material sheets for EMI shielding can be purchased in rolls anywhere," he said These materials have "piezoresistive" properties, which means that they change the resistance as they strain. Truby realized that if they were placed in some position on a moving object, they could be made into effective soft sensors. When the sensor is deformed in response to the extension and compression of the torso, its resistance will be converted to a specific output voltage, which will then be used as a signal related to the motion.
But this kind of material has low flexibility, which will limit its use in soft robot. Inspired by kirigami, Truby designed and used a laser to cut rectangular conductive silicon film into various patterns, such as rows of small holes or cross-sections similar to chain link fences. This makes them more flexible, stretchable, and "look beautiful," Truby said.
The body of the robot designed by the researchers consists of three parts, each with four fluid actuators (a total of 12) for moving the arm. They fused a sensor on each segment, and each sensor covered and collected data from an embedded actuator in the soft robot. They used "plasma bonding" technology, which allows the surface of one material to be energized and bonded to another. It will take about a few hours to form dozens of sensors that can be integrated into a soft robot using hand-held plasma bonding devices.
Image source: MIT
As assumed, in the experiment they put the sensor on a trunk, and the sensor did capture the overall movement of the trunk. But they are really noisy. "In essence, according to our traditional concept, they are not ideal sensors in many aspects, because noise is a very annoying thing in industry." Truby said. "But it's just a general fact that sensors are made of soft conductive materials. Higher performance and more reliable sensors require special tools that most robotics labs don't have. "
Therefore, in order to estimate the configuration of soft robot only by using sensors, researchers set up a deep neural network, and do most of the heavy work by screening noise to capture meaningful feedback signals. The researchers developed a new model to describe the shape of the soft robot in a kinematic way, thus greatly reducing the number of variables needed to process the model.
Image source: Ryan L. Truby, MIT CSAIL
In the comparative experiment, the researchers let the soft robot's trunk swing and extend itself for about an hour and a half with random configuration. They use the traditional motion capture system to obtain the real ground data. At the same time, in the training, the model also independently analyzes the data from its sensors to predict the configuration, and compares its prediction with the real ground data collected at the same time. In this way, the model "learns" to map signal patterns from its sensors to the actual configuration. The results show that for some more stable configurations, the estimated shape of the robot's sensors is consistent with the real situation on the ground.
Next, the researchers aim to explore new sensor designs to improve sensitivity, and develop new models and deep learning methods to reduce the training time and process required for each new soft robot. They also hope to improve the system to better capture the complete dynamic motion of the robot.
At present, the neural network and sensor skin of the soft robot are not sensitive to capture micro motion or dynamic motion. However, for the current learning based soft robot control method, this is an important first step, Truby said: "like our soft robot, human life system is not necessarily completely accurate. Therefore, compared with our human beings, robots are not exactly machines at the beginning. To make a robot that looks less accurate, we undoubtedly do a lot OK. "