fourier: "first-line customers don't need robots that can do somersaults, but they do need robots that can solve problems"
2024-09-26
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
on september 26, humanoid robot company shanghai fourier intelligent technology co., ltd. released a new generation of universal humanoid robot gr-2, which is 1.75 meters tall, weighs 63 kilograms, has 53 degrees of freedom, and can move a single arm with a load of 3 kilograms. the new generation of battery capacity has doubled, and the battery life has been extended to 2 hours. gu jie, founder of shanghai fourier intelligent technology co., ltd., told the paper that the popularity of humanoid robots has a positive effect on the industry, but he refuses to involuntarily roll back the industry. he hopes that people in the robot industry will be willing to spend 5, 10 or even 20 years to work hard and iterate step by step. at this stage, it is far from the time when the price of a humanoid robot will make a huge difference if it is 10,000 yuan more or less.
GR-2。
with 53 degrees of freedom throughout the body, the large model changes the paradigm of robot development
the gr-2 adopts the bionic design language of the previous generation gr-1 in appearance. it is 1.75 meters tall, weighs 63 kilograms, has 53 degrees of freedom, and a single arm motion load of 3 kilograms. the new generation of batteries has doubled the capacity, and the battery life is extended to 2 hours. the detachable design allows the robot to work for a long time. the overall layout adopts an internal wiring design, placing the cables inside the robot. the full series structure allows each joint to be independently calculated, controlled and predicted, which is convenient for debugging and maintenance and reduces manufacturing costs.
upper limbs are an important direction for gr-2 iteration. the gr-2 single-handed degrees of freedom have been increased from 6 in the first generation to 12, which can perform more anthropomorphic and complex movements. the dexterous hand is equipped with 6 array-type tactile sensors and motion control algorithms, which can sense the gripping force in real time, adjust the grasping strategy, and achieve precise force-controlled grasping of objects of different shapes, materials, and weights. gr-2 supports a variety of upper limb teaching methods such as vr remote control, drag-and-drop teaching, and host computer terminal control, quickly creates actions, and records multi-dimensional data such as upper limb motion trajectory, joint angle, strength, dexterous hand movement, and tactile information.
dexterous hands.
the actuator is the core component of the robot. seven new-generation fsa actuators will be used in key parts of the gr-2, such as the hip and ankle, to reduce the deadweight, increase the maximum joint torque by 30%, and enhance the anti-interference ability. the new version of the actuator uses a dual encoder system, which increases the encoder accuracy by 2 times, ensuring the control accuracy of the gr-2 when performing various actions.
talking about the future technological trends of humanoid robots, gu jie said that the realization of general robots requires the joint development of algorithms and ontology, and the embodied modal model requires the evolution of hardware carriers. the ontology layer includes full-body hardware designs such as lightweight integrated design and manufacturing, bionic torso design, and anti-drop impact design, as well as key components such as integrated actuators, dexterous hands, sensors, special chips, special power batteries, and controllers. at the algorithm level, the ai big model is the robot's brain, responsible for perception, interaction, and decision-making tasks. the ai small model is the robot's cerebellum, performing dual-arm collaborative planning and control, dual-dexterous hand dynamic grasping operations, dynamic full-body control, and intelligent behavior control.
"the future breakthrough of humanoid robots will definitely be the combination of vision, language and action. the robot itself will be able to run a large model, and the large model will be able to control the generalization of the robot's actions and complete common tasks, rather than being pre-programmed." gu jie said that if the robot can achieve this step, it will have a certain reasoning and generalization ability.
regarding the large-scale model landing robot, gu jie said that this is not just imagination. in the past six months, there have been documents showing that simple actions can be generated through large models, and robots can also learn actions through reinforcement learning. this is a paradigm shift in robot research and development. "in the past, to control more than 50 motors of a robot, you had to send instructions to it. if it is full-body motion control, every moment and every second is a different position, different force changes, and it is simply an impossible task to control through traditional programming. but now through large models, neural networks can control the generation of actions, and unexpected results will appear because the robot is self-taught."
customers don’t need robots that can do somersaults, and the industry is not at the stage of a price war
gu jie, who graduated from shanghai jiao tong university, started his first robot business in 2008. at that time, he wanted to develop a humanoid robot. considering the maturity of the technology, he temporarily gave up this direction and turned to rehabilitation robots. seven years later, the company was acquired. in 2015, he started another business and named it "fourier intelligence" after the mathematician fourier. in 2017, fourier released the first commercial lower limb exoskeleton robot in china. in july last year, the universal humanoid robot gr-1 was released, and 100 robots have been delivered so far.
gu jie admitted that neither gr-1 nor gr-2 humanoid robots have reached perfection, but only by going out of the laboratory can feedback be collected and iterative improvements be made. humanoid robots are still in the early stages of the industry, and there is a chicken-and-egg relationship between mass production and advanced technology. only by forming a flywheel of mass production, scene polishing, and customer feedback can the technology be truly put into practice.
"i have heard our front-line customers describe countless times that they do not need a robot that can do somersaults. what they really need is a robot that can stably and consistently complete tasks and solve problems every time it is started." zhou bin, vice president of fourier's general robot division, said frankly that motion performance is only one aspect of all dimensions of a robot. blindly improving motion performance will only affect the performance of other dimensions. the development of humanoid robots must be based on the final scenario and the final customer's comprehensive judgment, and compromises must be made in the trade-offs.
GR-2。
zhou bin believes that the development of humanoid robots will go through three stages: first, building a stable embodied intelligent body with superior performance and opening interfaces to achieve innovation. second, collecting massive training data in the physical world. third, using high computing power to achieve large-scale training and improve generalization capabilities.
at present, fourier has built a development interface solution based on mainstream programming languages, providing developers with a more open development platform, which helps robotics researchers, industrial automation engineers, ai software developers, etc. to efficiently customize diversified applications and realize robotics technology innovation. the development platform supports open source frameworks such as nvidia isaac lab, ros, mujoco, and webots.
competition in the humanoid robot industry is fierce. gu jie said that it has been 22 years since he began designing robots as a student in 2002. more and more teams have joined the tortuous, difficult, but hopeful road of humanoid robots, and a large amount of capital and talent have poured in. however, the development of robots is not achieved overnight. this is a long-distance industry. perhaps the development speed this year is faster, but it does not mean that this year is the moment of the decisive battle. humanoid robots are far from the stage of price wars, and fourier does not engage in price wars. "we pay more attention to making the products mature and truly generating commercial value, rather than just caring about the price now. of course, if a robot costs several million or tens of millions of yuan in the end, no one can afford it, but now is far from the time when there will be a huge difference if the price is 10,000 yuan more or 10,000 yuan less."
he said that the industry's boom has a positive effect, but rejected ineffective involution. "we hope that all practitioners in this industry truly believe that robotics is a long-term career and are willing to spend 5, 10, or 20 years to work on it, iterating step by step."
the paper reporter zhang jing
(this article is from the paper. for more original information, please download the "the paper" app)
