news

the difficulty is that musk wants to build people, not machines.

text丨he qianming and li zinan
editor: he qianming and huang junjie

"even if i am abducted by aliens tomorrow, tesla will solve the problem of autonomous driving. but i don't know how tesla will win with optimus." at the earnings conference in april this year, musk gave a reason why he should receive a salary of $50 billion.

a person who has been in contact with the tesla team told us that shortly before musk said this, he asked the humanoid robot team to come up with a meaningful breakthrough within this year - to allow robots to go on the production line to complete some work. if this is not achieved, further layoffs will be made, and "the team is under increasing pressure."

we learned that hundreds of optimus humanoid robots have been produced, but they are not yet ready to work on the production line. they can only be used to collect data and test different design solutions.

this humanoid robot project, which began in 2021, is completely opposite to the logic of tesla's development of other products. in car manufacturing, tesla started with the highly profitable electric sports car roadster, gradually improved production efficiency, and built the cheap model 3. autonomous driving is also from low to high, and it is improved while selling.

as for humanoid robots, musk wants to achieve mass production in one step at the ultimate form: the cost should be as low as less than $20,000 - lower than the annual income of an american blue-collar worker; but smart enough to replace some workers' jobs.

in the past three years, musk has talked about optimus as often as autonomous driving and robotaxi in public, and sees it as the hope to support tesla's valuation. "if we have humanoid robots with perception capabilities that can perform tasks as required, productivity will increase significantly and there will be no limit to economic growth."

china's stock market has hyped up several rounds of tesla's robotics supply chain concept stocks, but musk's grand story has done little to boost tesla's stock price. since its high in 2022,teslathe market value of the company has fallen by nearly 40%.

the challenge facing the tesla robotics team is that it is basically impossible to create a robot that can act like a human in one step. at present, they can only make some engineering compromises, perhaps targeting specific work scenarios, such as screwing, and specially training models to allow humanoid robots to play a part first. as for whether musk can accept the result of the compromise, it is still unknown.

the biggest difficulty: creating a human being, not a humanoid robot

according to the biography of elon musk, in early 2021, musk played a video of boston dynamics' humanoid robot atlas at a tesla executive meeting. in the video, the humanoid robot can perform difficult movements flexibly like a human. he told the executives attending the meeting, "whether you like it or not, humanoid robots will appear. we should get in on the game."

his goal is clear: it must be a humanoid robot with a human-like body shape, not a four-legged or wheeled mechanical device made by boston dynamics or other companies. and the robot must learn by visual observation like a human, and learn to complete the work by hand, instead of having engineers input every line of work instructions in advance like boston dynamics or traditional robotic arm companies.

boston dynamics released a video of the atlas robot at the end of 2020. source: boston dynamics.

previous robots were designed for specific scenarios, which resulted in limited versatility and production capacity. according to data from the international federation of robotics, the total sales of industrial robots worldwide in 2023 will be only 590,000 units, about half the production of rolex watches. these 590,000 units are at least dozens of specialized forms that are not interchangeable with each other, and most of them are produced in a few hundred units per year, with no scale effect at all.

now all humanoid robot companies are telling the story of tesla robots: our workplaces and tools are designed for people. if humanoid robots become a reality, they can complete multiple tasks with one set of hardware without the need for custom development, and mass production of the same model will be possible. when musk talks about the humanoid robot project, he often talks about scale, saying that he wants to produce millions, tens of millions, or even billions of units. "mass production is also important for humanoid robots, so that they are not so expensive."

the difference is that few humanoid robot companies insist on making humanoid robots. they usually choose a compromise route, using various clamps to replace human hands or using wheels to replace legs, first solving some problems and generating some commercial value, and then iterating towards humanoid form step by step.

but musk did not compromise. in weekly discussions on humanoid robots, he rejected a large number of designs that did not look like humans, such as installing different tools on the robot's arms to perform different tasks. he believed that only hands were enough.

tesla engineers also felt that the robot's pinky finger was useless, but because it was "too scary" and too unlike a human to remove it, it was retained. the alternative was to extend the length of the pinky finger so that it could play a greater role. musk also very specifically required that the robot's fingers have the taper of female fingers. he set the robot's height to 5 feet 8 inches (nearly 1.73 meters), slightly shorter than the average height of an adult male in the united states; he required the robot to look like an androgynous elf so that "you wouldn't feel that it might hurt you."

an investor who has been in contact with the tesla team said that musk's insistence on humanoids was to replace the last link that restricts automobile manufacturing - workers. at that time, tesla had completely escaped the production capacity hell, sold 500,000 cars last year, and the continued rise in stock prices pushed musk to become the world's richest man. but tesla's car factory in california produces 8,550 cars per week, approaching the production capacity limit. tesla realized that the efficiency of workers has reached its limit. if it wants to be faster, it can only build tools that are stronger than humans.

automobile manufacturing is one of the most automated industrial production processes. the first use of robotic arms, which were invented in the 1960s, was to move and weld parts in automobile factories. in tesla's automobile factory, only the final assembly process requires a large number of workers to move around the workstation in a small range with their legs, assemble parts with their hands, or tighten screws with electric drills. the most fundamental way to replace them is to create similar people.

if a robot can "understand" how to work through machine learning, walk like a human, and raise its hands to complete the work in the tesla car factory, then it should also be able to be used in other jobs and produce economies of scale. this is a natural logical deduction, but each of the premises is difficult. after three years, the tesla team has not even a clear path to achieve it.

fuzzy exploration: how to use mechanical parts to simulate humans

before tesla, scientists around the world had been working on creating a usable humanoid robot for nearly 50 years. the task they faced was easy to understand from the beginning to the end: use metal, plastic, silicon and other materials to simulate human bones and flesh, and create a machine that can think and move like a human.

the most famous humanoid robot in the industry is atlas, made by boston dynamics. it can do backflips, run fast on uneven grass, and dance flexibly indoors. but boston dynamics gave up on making the two most unique organs of humans - hands and brains - from the beginning. atlas can perform difficult movements, but only limited to pre-set movements, and the ends of its arms are round bumps like doraemon.

when tesla was developing a humanoid robot, the first thing it did was the human hand. the human hand has 27 degrees of freedom, is flexible in movement, is covered with highly sensitive tactile nerves, and has strong force control capabilities.

in the early plan, tesla installed 6 drive motors in the robot hand, allowing optimus to hold most objects, but it was far from flexible. among the five fingers, only the thumb was equipped with 2 motors, which could bend inward and swing left and right, and the remaining 4 fingers were only equipped with a single motor that could control the bending and swinging. each drive of the hand has a metal wire connected to the motor, and the fingers bend and stretch accordingly when the motor retracts and releases the metal wire.

the second-generation optimus hand currently has only 11 degrees of freedom. in may this year, musk claimed that the optimus hand's degrees of freedom would be increased to 22 within this year, closer to the human hand. now tesla has a plan for a new generation of robotic hands.

but human hands are made of flesh and blood, not welded or riveted together like robotic hands. when a human hand moves, the wrist will not collide with the palm in front, and power will not be lost inside the human hand, but these problems may occur in a robotic hand. the integrity of the robotic hand structure is related to the stability and continuity of the hand movement. tesla removed many of the linkage devices in the optimus hand and used a cable drive made of cables to imitate the muscle tissue of the human hand, hoping to minimize the power loss and hysteresis generated when the robotic hand moves.

people perceive pressure through limb deformation and the compression of nerve endings, while optimus measures pressure through the current of the actuator. the actuator provides optimus with the power to hold the object, and the current corresponds to the output force. the human hand has tens of thousands of tactile sensors, while optimus's hand only has 6 pressure sensors and cannot sense temperature.

tesla uses some linear actuators in the leg joints. compared with traditional rotary actuators, linear actuators can be arranged longitudinally to maximize the use of the internal space of the legs, arrange larger and longer motors, and provide greater propulsion. the leg actuators of the second-generation optimus are strong enough to lift a half-ton piano. but the disadvantage of this is that the balance of the robot's legs is more difficult to control, and the difficulty of developing control algorithms has increased exponentially.

before letting optimus walk, tesla also measured the transmission mechanism and torque of each joint of the human leg when walking, so that optimus's movements could be as human-like as possible. structurally, the human knee is an efficient four-link device. the four bones plus the knee joint can perform actions such as walking, jumping, squatting, etc., and accurately control the center of gravity to keep people balanced and not fall. optimus can only walk slowly at present, and cannot successfully complete actions such as squatting.

the human body has more than 200 degrees of freedom, which are connected by hundreds of bones and dozens of joints. tesla has simplified it to 28 degrees of freedom in the optimus trunk. these degrees of freedom are realized by three different sizes of rotary actuators and three linear actuators. the rotary actuators are distributed in the shoulders and hips that need large angles of rotation, and the linear actuators are distributed in the knees, elbows and other parts with small swing angles.

each actuator is equipped with components such as lead screw, reducer, motor, sensor and encoder. the motor provides power for the actuator, the lead screw converts the force of component rotation into tensile force to control the contraction and extension of the joint; the reducer acts like the connection between human bones and joints, and can amplify the force through leverage.

improving body freedom is only the first step in using optimus to simulate humans. tesla is still continuing to reduce the weight of optimus, which allows it to complete more tasks with the same power. the second-generation optimus weighs 56.6 kg, which is lighter than an adult human.

to make a good humanoid robot, flexible, durable, stable and cost-controlled hardware components and materials are just the foundation. for example, people can catch moving objects because the human brain has cognition of physical laws such as gravity and acceleration, so it can predict the trajectory, and the body has the ability to move, so it can complete this seemingly simple action.

in 2022, tesla transplanted the machine vision neural network in the fsd (full self-driving) on ​​the car to optimus. just like teaching a car to drive, optimus can understand the real world and respond appropriately, allowing it to learn and improve from human actions.

"autonomous driving requires cars to avoid colliding with any objects in a 2d world, but the robot's task is to actively collide and touch an infinite number of things in the 3d world." an investor who has been in contact with the tesla team said that the technical complexity of the entire autonomous driving process is far less than that of a robot's hand.

"the bigger problem is that there are no scaling laws in the robotics field." an investor in the robotics field said that this is a difficult problem faced by the entire intelligent robotics industry. on the one hand, they cannot find enough data to train robot models, and on the other hand, there is no suitable model architecture that can digest massive amounts of data.

tesla began recruiting humans to be optimus's teachers last year, with a clear requirement of 170cm-175cm in height, similar to the robot. these people have to wear motion capture devices with sensors all over their bodies to demonstrate to the robot how people use their joints to move. some analysts speculate that tesla may extract key data on human movement from massive videos in the future and feed it to optimus, which may speed up its learning.

according to the current progress, humanoid robots need a long period of learning and training to meet musk's minimum requirements and replace some workers in tesla's car factories. a tesla analyst said that by the middle of this year, optimus's pass rate in walking tests of more than 500 meters was less than 60%.

its competitors are hourly workers in texas, usa, who earn $22 per hour. these workers can basically complete a series of complex tasks within 50 seconds, such as picking up parts, finding the right position to complete assembly, and reviewing quality. these people will not fall down and wait for help after walking 500 meters. they can complete complex communication and collaboration with others, but they are far from the most efficient and lowest-paid workers in the world.

musk can currently only point out two aspects of optimus that are superior to humans: it is more serious and does not slack off than humans; and it can work continuously for 16 hours when fully charged.

advanced preparation: consider mass production at the concept stage

"(humanoid robots) are extremely difficult to mass-produce and are underestimated." when ai day was held in august 2021, tesla's humanoid robot project was still a concept, and musk foreshadowed the difficulties that would be faced.

when designing humanoid robots, tesla's team was looking for ways to reduce the cost and difficulty of mass production. according to the biography of musk, a typical example is how to make the robot hand control the pressure applied and feel the pressure.

engineers have proposed a variety of solutions, such as using the current of actuators in the knuckles to measure pressure, placing capacitors on the fingertips similar to those used on mobile phone touch screens, or embedding air pressure sensors or chips and placing a miniature camera.

the common practice in the robotics industry is to install pressure sensors in the joints of the human hand, which is more expensive and difficult to manufacture.

seeing these plans, tesla's chief designer von holzhausen only asked one question: what is the difference in cost? the final solution is to use the current in the joints, because it does not increase the number of parts - the cost is controllable and easy to produce.

tesla's experience and resources in car manufacturing have also been transferred to the humanoid robot project. the motors, batteries, software systems, etc. in the humanoid robot were all modified from similar parts in electric vehicles in the early versions, and will be optimized using the software infrastructure that tesla has built over the years.

just like analyzing how motors drive electric vehicles through the city, they used a simulation system to simulate the performance of a humanoid robot's 28 joints in performing a variety of actions, identify the commonalities, and simplify them into six solutions with lower unit cost and weight.

there are many obvious problems with optimus at this stage. people close to tesla said that the software team is still trying to find a way to use code to replace the two six-axis force sensors on optimus' ankles, which are bulky and expensive and only provide partial redundancy for the robot's stability.

for tesla, the biggest challenge in mass-producing humanoid robots is to make them useful and good enough to use. tesla has not yet built a robot assembly line, and can only assemble them manually. the cost of a single robot is about $100,000, but the value it can create now is minimal.

once tesla has developed sufficiently useful robots, it will have to build a low-cost supply chain from scratch.

although the industry has experience in producing various parts used in humanoid robots, they are usually used in high-end manufacturing scenarios and are expensive. for example, ball screws are generally used in expensive heavy machinery or aerospace equipment. tesla made advance preparations. before the robot design was finalized, the team began to build a low-cost supply chain.

since mid-2022, tesla has started to ask chinese parts suppliers to send parts for testing. among them, the more representative ones are tesla electric vehicle suppliers top and sanhua intelligent control, which assemble actuators. they are tesla's long-term suppliers, supplying cast aluminum structural parts for the car. specifically at the component level, there are hollow cup motors used in robot hands sent by mingzhi electric to tesla, and harmonic reducers used in joint actuators sent by greenland harmonic.

there are still many uncertainties as to whether these parts suppliers can eventually enter tesla's robot industry chain. cost is the advantage of chinese suppliers. an analyst in the field of robotics said that for the same parts, the price of chinese companies is at least 30% cheaper than that of well-known overseas companies, and some are even 80% cheaper. however, there is still a big gap in performance between the two sides. we understand that in the first batch of samples sent in the c round starting in the middle of this year, basically no chinese parts suppliers passed tesla's requirements.

an investor who has been in contact with the tesla team said that tesla's internal goal is to establish a humanoid robot supply chain outside of china. this means that if chinese companies want to supply tesla robots, they need to invest in building factories overseas first.

tesla assembled a batch of humanoid robots for testing at its austin factory in the united states. as of june this year, only two of them were moving batteries in the factory. there is no value in using a humanoid robot that costs more than $100,000 to move lithium batteries. battery factories generally use unmanned mobile carts with pallets that cost less than $10,000 to load batteries, and can move thousands of batteries at a time.

musk once again started his usual cycle: using unrealistic timetables to push team members beyond their limits and approaching the goal in repeated delays. however, sometimes, the gap between expectations and reality is so large that it cannot be filled in one or two generations, such as the electric car promoted by edison a hundred years ago.

title image: visual china