Cutting-edge technology and innovative application of humanoid robots
Humanoid robots have a history of 100 years, and they have been produced and disappeared in the process, which is summarized as two heavens of ice and fire. Based on forty years of control experience, based on bipedal robots replaced for four generations, from the earliest spring inverted plate, to model pre-control 1.0, 2.0, 4.0. What we have seen today is more wonderful performances, and behind the scenes is a lot of training behind the scenes.
There are many genres and methods for quadruped and bipedal robots, and the mainstream worldwide includes inverted pendulums, spring inverted pendulums, dynamic models and model pre-control, etc. or a combination of these methods.
All Atlas performances do not exceed 15 minutes because the power does not last longer. Atlas has a new action appeared, the middle months are constantly rapid iteration, the iteration speed is not fast, because the current based on dynamic traditional control, the first problem is through a large number of physical space iteration, the speed is slow, the second problem is that all Atlas actions are linear, driven by dynamics, each of his movements must be very exquisite, so he can make such clear actions, his requirements for many components are very high, our country can not make Atlas because there are not enough parts, The reducer capacity is not enough, and the small part technology is limited, but Atlas also has a long way to go to application.
We came to the conclusion that it turns out that it is difficult for robots to go to industry and industry through very strong dynamic characteristics that are coupled with software and hardware, so how to explain it? That is, now the dynamics of dimensionality reduction, dimensionality reduction of a robot many joints, the dynamics of its whole body, if this mode, or degrees of freedom, dozens of hundreds, but we often abstract it into a two, three, inverted plate, spring inverted plate, that is to say, we reduce the dynamics dimensionality, and then we have to make its joints very fine, equivalent to us through super hardware, hardware includes calculation, including drive, including execution, we can make a movement, this robot movementA very sophisticated robot can complete the action of a robotic arm on the production line, but we let it run and jump and perform a lot of actions, which is a very difficult task. So how do we crack this problem?
There are two ways, one is to control dimensionality reduction, this hardware overconfiguration is what we are doing at present, we continue to make this joint and actuator of the robot and its control more and more exquisite, the second is to assemble its software, another way to solve this problem.
In the past ten years, whether it is Aslas, including the domestic model of our four-legged and bipedal robots, basically software engineers have been waiting, hardware engineers have been repairing, he has not decoupled, he must move the robot before debugging, so this iteration of ours is very slow. The second problem is that it is very expensive now, and the most expensive components are line-line actuators and rotary actuators. How to reduce costs, this robot must be tens of thousands of dollars, hundreds of thousands of dollars, hundreds of thousands of dollars, then we are using the traditional industrial chain, through manufacturing or through additive manufacturing of some new processes, this is a path, the second is industrial technology reengineering, recreate a new industry.
Some considerations on these issues, the practice of quadruped robots, we now have 13 years of experience. Over the course of these years, we have mainly done two things,
One is to build the core components, one is to build our control model, now we do humanoid robots, now with the support of the local government, we have made small robot dogs, through these years, we have created the first high-explosive high-torque robot joint machine, and foot robot operating system, we first reach a US-level control. To see whether a robot dog is good or not, it depends on its agility, and over the years, the components and software of the core technology have been created on the quadruped robot, and the software includes the operating system, the cloud platform behind it, and the algorithm library.
There is also a problem here, all these things are our own creation, the United States is engaged in Accras, sports mini, many key components can be bought from Amazon, such as Muge's valve, including Honeyface sensors, but because many of them are restricted to our country, in the future we will do robot dogs, robots, we should be able to buy parts from Jingdong Taobao, the 20 major reports also put forward, we need to implement industrial technology reengineering projects, 5G industrial technology reengineering is not along the traditional industrial chain thinking, through large-scale production to reduce costs, we are to create a new ecology, a new model through AI-driven. Our humanoid robot is benchmarked against general artificial intelligence, a general humanoid robot industry chain, not only the previous industrial robot reducer, motor, actuator,
The point is to use AI, a new one empowered by intelligence. Traditional industrial robots they master the motor, reducer, controller this core component, to the leg and foot robot, quadruped as the representative, it has developed to the electric joint, walking control and perception head, and then to the humanoid robot industry chain, its focus is on high-end chips, supercomputing centers, operating systems. Now is in the era of modularization sensed by the small total library of electrical joints, so the future to do humanoid robots is to enter the era of generalization with computing supercomputing networking, humanoid robots face a new ecology, a new industrial chain model, so how to solve this problem, there are many ways, the current need to do is very important to solve this problem through software and hardware coupling.
The traditional way of iteration is too slow, everyone loses patience, capital can not be rewarded, and can not go to the industry, so this model is to improve it, otherwise it is very difficult to go down, now there are also successful cases, Zurich Swiss Institute of Technology anymore and American digital robot, it is transplanted the neural network to the robot, and then transplanted to the quadruped or bipedal, to achieve the decoupling of software and hardware, through the neural network, machine learning driven this robot into a rapid iteration.
So now it is necessary to take the neural network like the control of the robot's state quantity, and then reward the dynamics with the maximum reward. To put it simply, we went from the earliest single spring inverted pendulum simple model, later to the model and control dynamic model, now to the era of data-driven artificial intelligence, now the machine learning neural network is transplanted to the robot, now the CV model is more mature, can do monitoring, but artificial intelligence to change our industry, industrial ecology There are not many examples, so it is very difficult to transplant the neural network to a quadruped robot.
How to do it specifically, we first use the model based on this system to identify CAD to build a robot dog model, through pre-training knowledge distillation, digital space training to transplant it into the robot dog's controller, and then the original dynamic model for comparison, and then superimposed on the robot dog for training, this effect is very obvious. After linking the neural network with the robot dog dynamics, the movement speed is very fast, and now we are to transplant this technology into this next humanoid robot development plan, the following is what we used to do called lifelong learning, about the cerebellum of kinesiology, what is lifelong learning?
That is, people to a certain extent, in middle age, the ability will decay will not be stronger, the ability of the robot dog has been improving, the AI generation of action behavior of this brain, to it, will accelerate the development of humanoid machines. Now we input the language image behavior to let him do what, and then from a large amount of data to our behavior, by continuing the training process, the iteration speed is very fast. With this theory, the next step is to do a good job with technology, this is a mutual relationship, now a high-level performance of general artificial intelligence is to gather intelligence, can sense, can think, and can act, industrial intelligence empowers to general humanoid robots, they interact,
The first role is to quickly make humanoid robots move, train models to drive robots through intelligent computing power in the cloud background, more by training, through this model to make humanoid robots mature rapidly, and then another is to transform a very good application prospect for general artificial intelligence, which can now empower industry. Now the next step is to do a new generation of humanoid robots, which is the hardware of the general-purpose zero machine plus the general-purpose artificial intelligence plus the GPU supercomputing center.
The future robot will be guided by strong artificial intelligence, and there will be quantum computing in the background, truly realizing human-machine co-prosperity, and realizing the development that unimaginable humanoid robots can bring to our industry and society. With technology in humanoid robots through sensors, power and actuators, in through artificial intelligence and technology is a giant model operating system and supercomputing center, so to do a good job of these robots, these real technologies to combine them, just like to engage in intelligent networked cars, now we must engage in software-defined cars, software-defined cars are to solidify hardware, standardize, and then iterate through software to achieve new work. Whether it is a traditional force to build a car, or an Internet company to build a car, it must take the road of software-defined car, so that the car can have competitiveness and development. We now want to decouple the software and hardware, and also empower artificial intelligence technology to the software functions of humanoid robots, so the software and hardware structure we do, and then the intelligence to empower our functions, then we are equivalent to walking out of this future with more advanced capabilities than our intelligent networked cars.